| 1 | """v3.0.0 orchestration pipeline.""" |
| 2 | |
| 3 | from __future__ import annotations |
| 4 | |
| 5 | import contextlib |
| 6 | import copy |
| 7 | from collections.abc import Iterable, Mapping |
| 8 | import math |
| 9 | import queue |
| 10 | import re |
| 11 | import sqlite3 |
| 12 | import sys |
| 13 | import threading |
| 14 | import time |
| 15 | from collections import Counter |
| 16 | from concurrent.futures import ThreadPoolExecutor, as_completed |
| 17 | from dataclasses import dataclass, field, replace |
| 18 | from datetime import date, datetime, timedelta, timezone |
| 19 | from pathlib import Path |
| 20 | from shutil import which |
| 21 | from typing import Any |
| 22 | |
| 23 | from . import ( |
| 24 | amazon, |
| 25 | arxiv, |
| 26 | bird_x, |
| 27 | bluesky, |
| 28 | brightdata, |
| 29 | corpus, |
| 30 | dates, |
| 31 | dedupe, |
| 32 | digg, |
| 33 | dripstack, |
| 34 | entity_extract, |
| 35 | env, |
| 36 | github, |
| 37 | grok_x, |
| 38 | grounding, |
| 39 | hackernews, |
| 40 | health, |
| 41 | hiring_signals, |
| 42 | http, |
| 43 | instagram, |
| 44 | jobs, |
| 45 | linkedin, |
| 46 | library, |
| 47 | library_index, |
| 48 | log, |
| 49 | meta_ads, |
| 50 | normalize, |
| 51 | permission_preflight, |
| 52 | perplexity, |
| 53 | pinterest, |
| 54 | planner, |
| 55 | polymarket, |
| 56 | providers, |
| 57 | query, |
| 58 | reddit, |
| 59 | reddit_listing, |
| 60 | reddit_public, |
| 61 | relevance, |
| 62 | rerank, |
| 63 | schema, |
| 64 | signals, |
| 65 | snippet, |
| 66 | stocktwits, |
| 67 | techmeme, |
| 68 | telegram, |
| 69 | threads, |
| 70 | tiktok, |
| 71 | topic_shape, |
| 72 | truthsocial, |
| 73 | trustpilot, |
| 74 | x_api, |
| 75 | x_envelope, |
| 76 | x_judge, |
| 77 | xai_x, |
| 78 | xiaohongshu_api, |
| 79 | xquik, |
| 80 | xurl_x, |
| 81 | youtube_yt, |
| 82 | ) |
| 83 | from .cluster import cluster_candidates |
| 84 | from . import fusion |
| 85 | from . import render |
| 86 | from .fusion import collapse_duplicate_urls, weighted_rrf |
| 87 | |
| 88 | DISCOVERY_SOURCES = ("reddit", "hackernews", "digg", "x") |
| 89 | _DISCOVERY_GENERIC_DOMAIN_TERMS = { |
| 90 | "ai", "artificial", "intelligence", "tech", "technology", "trending", "trend", |
| 91 | } |
| 92 | |
| 93 | DEPTH_SETTINGS = { |
| 94 | "quick": {"per_stream_limit": 6, "pool_limit": 15, "rerank_limit": 12}, |
| 95 | "default": {"per_stream_limit": 12, "pool_limit": 40, "rerank_limit": 40}, |
| 96 | "deep": {"per_stream_limit": 20, "pool_limit": 60, "rerank_limit": 60}, |
| 97 | } |
| 98 | |
| 99 | SEARCH_ALIAS = { |
| 100 | "hn": "hackernews", |
| 101 | "bsky": "bluesky", |
| 102 | "truth": "truthsocial", |
| 103 | "web": "grounding", |
| 104 | "xhs": "xiaohongshu", |
| 105 | "meta": "meta_ads", |
| 106 | "meta-ads": "meta_ads", |
| 107 | "xquik": "x", # xquik is a backend of the single "x" source, not its own source |
| 108 | } |
| 109 | |
| 110 | # trustpilot is capped at 1: every subquery would use the identical company |
| 111 | # identifier, so N streams are pure redundancy -- and each extra stream risks |
| 112 | # its own WAF-cookie Chrome harvest. |
| 113 | # amazon is capped at 1 for the same reason as trustpilot: the model supplies |
| 114 | # one product keyword for the run, so every subquery would issue the identical |
| 115 | # product search. Extra streams would be pure redundancy at one credit each. |
| 116 | # meta_ads is capped at 1 for the same reason as amazon: one advertiser page is |
| 117 | # resolved per run, so every subquery would issue the identical page fetch. |
| 118 | MAX_SOURCE_FETCHES: dict[str, int] = { |
| 119 | "x": 2, "jobs": 1, "linkedin": 1, "stocktwits": 1, "trustpilot": 1, "amazon": 1, |
| 120 | "telegram": 1, "meta_ads": 1, |
| 121 | } |
| 122 | |
| 123 | # Sources whose thin result is their normal success state, so the "<3 items" |
| 124 | # retry would re-fetch them after every success -- bypassing |
| 125 | # MAX_SOURCE_FETCHES and, for the resolved-entity sources, re-resolving |
| 126 | # WITHOUT the caller's override (a lookalike-misattribution path). |
| 127 | # trustpilot returns at most ONE item by design. |
| 128 | # perplexity answers once per run. |
| 129 | # meta_ads resolves one advertiser page per run, so a brand that genuinely |
| 130 | # ran two creatives this month is complete; a retry would re-resolve the |
| 131 | # page and re-spend the discovery credit. |
| 132 | THIN_RETRY_EXEMPT: frozenset[str] = frozenset({"trustpilot", "perplexity", "meta_ads"}) |
| 133 | |
| 134 | # Stream-artifact keys promoted to named top-level report artifacts. A stream |
| 135 | # artifact only ever reaches the report as an anonymous entry in the grounding |
| 136 | # list, so anything the renderer needs by name has to be lifted out of it -- |
| 137 | # most importantly on a zero-item run, which is exactly when naming the |
| 138 | # resolved advertiser and its counts matters most. |
| 139 | STREAM_ARTIFACT_LIFT_KEYS: tuple[str, ...] = ("meta_ads_page", "meta_ads_tally") |
| 140 | |
| 141 | |
| 142 | def _lift_stream_artifacts(bundle) -> None: |
| 143 | """Promote per-stream artifacts the renderer reads by name.""" |
| 144 | for stream_artifact in bundle.artifacts.get("grounding", []): |
| 145 | if not isinstance(stream_artifact, dict): |
| 146 | continue |
| 147 | for key in STREAM_ARTIFACT_LIFT_KEYS: |
| 148 | value = stream_artifact.get(key) |
| 149 | if value: |
| 150 | bundle.artifacts[key] = value |
| 151 | |
| 152 | |
| 153 | _FAILURE_SPECIFICITY = { |
| 154 | health.AUTH_FAILED: 0, |
| 155 | health.PAYMENT_REQUIRED: 1, |
| 156 | health.RATE_LIMITED: 2, |
| 157 | health.SCHEMA_DRIFT: 3, |
| 158 | health.TIMEOUT: 4, |
| 159 | health.UNREACHABLE: 5, |
| 160 | health.ERROR: 6, |
| 161 | } |
| 162 | |
| 163 | |
| 164 | @dataclass |
| 165 | class PaidSourceBudget: |
| 166 | """Command-wide, thread-safe budget for paid source adapter calls.""" |
| 167 | |
| 168 | used: int = 0 |
| 169 | owner: str | None = None |
| 170 | _lock: Any = field(default_factory=threading.Lock, repr=False) |
| 171 | |
| 172 | def try_consume(self, limit: int, *, claimant: str | None = None) -> bool: |
| 173 | with self._lock: |
| 174 | if self.owner is not None and claimant != self.owner: |
| 175 | return False |
| 176 | if self.used >= limit: |
| 177 | return False |
| 178 | self.used += 1 |
| 179 | return True |
| 180 | |
| 181 | |
| 182 | def _source_fetch_cap(source: str, config: dict[str, Any]) -> int | None: |
| 183 | """Return the effective per-run cap for one source. |
| 184 | |
| 185 | Every Perplexity adapter call is paid, and ``both`` performs two paid POSTs. |
| 186 | A generic fetch-cap override must not multiply either normal or Deep |
| 187 | Research mode across planner subqueries. |
| 188 | """ |
| 189 | override = config.get("_max_source_fetches") |
| 190 | if source == "perplexity": |
| 191 | return 1 if override is None else min(1, int(override)) |
| 192 | cap = MAX_SOURCE_FETCHES.get(source) |
| 193 | if cap is not None and override is not None: |
| 194 | return int(override) |
| 195 | return cap |
| 196 | |
| 197 | |
| 198 | def _resolve_depth_settings(depth: str, config: dict[str, Any]) -> dict[str, int]: |
| 199 | """Depth profile with optional CLI cap overrides applied (issue #716). |
| 200 | |
| 201 | Returns a copy so the module-level DEPTH_SETTINGS is never mutated. Overrides |
| 202 | are set directly (not max()) so callers can also lower a cap. `--max-results` |
| 203 | raises the final ranked pool (pool_limit/rerank_limit); `--max-per-source` |
| 204 | raises the per-stream truncation applied before pooling. The per-source fetch |
| 205 | cap (`--max-source-fetches`) is applied separately at the fetch site. |
| 206 | """ |
| 207 | settings = dict(DEPTH_SETTINGS[depth]) |
| 208 | # `is not None` (not truthiness) so an explicit 0 is honored as a real lower |
| 209 | # bound rather than ignored as "unset" — matches how main() stashes these. |
| 210 | max_per_source = config.get("_max_per_source") |
| 211 | if max_per_source is not None: |
| 212 | settings["per_stream_limit"] = int(max_per_source) |
| 213 | max_results = config.get("_max_results") |
| 214 | if max_results is not None: |
| 215 | settings["pool_limit"] = int(max_results) |
| 216 | settings["rerank_limit"] = int(max_results) |
| 217 | return settings |
| 218 | |
| 219 | # Per-handle result caps for the X handle-search lanes. The FROM lane (the |
| 220 | # subject's own timeline) is the single best source for a person topic, so it |
| 221 | # gets the highest cap; the ABOUT (mention) and related-handle lanes stay |
| 222 | # modest so total volume and request budget don't balloon. |
| 223 | FROM_LANE_COUNT_PER = 8 |
| 224 | MENTION_LANE_COUNT_PER = 5 |
| 225 | RELATED_HANDLE_COUNT_PER = 3 |
| 226 | |
| 227 | |
| 228 | def _has_perplexity_provider(config: dict[str, Any]) -> bool: |
| 229 | # Prefer direct Agent/Search APIs, but preserve the synchronous OpenRouter |
| 230 | # Sonar fallback for existing installs. |
| 231 | return bool( |
| 232 | config.get("PERPLEXITY_API_KEY") or config.get("OPENROUTER_API_KEY") |
| 233 | ) |
| 234 | |
| 235 | MOCK_AVAILABLE_SOURCES = [ |
| 236 | "reddit", |
| 237 | "x", |
| 238 | "youtube", |
| 239 | "tiktok", |
| 240 | "instagram", |
| 241 | "hackernews", |
| 242 | "bluesky", |
| 243 | "truthsocial", |
| 244 | "polymarket", |
| 245 | "grounding", |
| 246 | "xiaohongshu", |
| 247 | "github", |
| 248 | "perplexity", |
| 249 | "threads", |
| 250 | "pinterest", |
| 251 | "digg", |
| 252 | "arxiv", |
| 253 | "techmeme", |
| 254 | "trustpilot", |
| 255 | "amazon", |
| 256 | "meta_ads", |
| 257 | "jobs", |
| 258 | "linkedin", |
| 259 | "corpus", |
| 260 | "dripstack", |
| 261 | "telegram", |
| 262 | ] |
| 263 | |
| 264 | |
| 265 | def normalize_requested_sources(sources: list[str] | None) -> list[str] | None: |
| 266 | if not sources: |
| 267 | return None |
| 268 | normalized = [] |
| 269 | for source in sources: |
| 270 | key = SEARCH_ALIAS.get(source.lower(), source.lower()) |
| 271 | if key not in normalized: |
| 272 | normalized.append(key) |
| 273 | return normalized |
| 274 | |
| 275 | |
| 276 | def available_sources( |
| 277 | config: dict[str, Any], |
| 278 | requested_sources: list[str] | None = None, |
| 279 | *, |
| 280 | x_pending: bool | None = None, |
| 281 | local_only: bool = False, |
| 282 | x_envelope: bool = False, |
| 283 | suppress_x_host_lane: bool = False, |
| 284 | ) -> list[str]: |
| 285 | """List the sources the next run can serve. |
| 286 | |
| 287 | ``local_only=True`` is the safe/diagnose flavor (doctor's permission |
| 288 | block): availability is answered from local evidence only, so the X |
| 289 | check never spawns xurl's live ``whoami`` network call. Research-time |
| 290 | callers keep the default live semantics. |
| 291 | |
| 292 | X is listed when an engine backend is available, or browser auth is |
| 293 | pending, or the hosting model declared the X connector lane |
| 294 | (``env.x_host_lane_declared``), or a validated ``--x-posts`` envelope is |
| 295 | present for this run (``x_envelope``), in every cookie mode. |
| 296 | ``suppress_x_host_lane`` turns only the lane branch off (discovery |
| 297 | enrichment passes); an envelope still counts. |
| 298 | """ |
| 299 | available: list[str] = [] |
| 300 | # reddit_public needs no API key - always available |
| 301 | available.append("reddit") |
| 302 | if corpus.resolve_directories( |
| 303 | config.get("_CORPUS_DIRS"), config.get("LAST30DAYS_CORPUS_DIRS") |
| 304 | ): |
| 305 | available.append("corpus") |
| 306 | if config.get("SCRAPECREATORS_API_KEY"): |
| 307 | available.extend(["tiktok", "instagram"]) |
| 308 | if env.get_x_source(config, local_only=local_only): |
| 309 | available.append("x") |
| 310 | elif x_envelope or ( |
| 311 | not suppress_x_host_lane and env.x_host_lane_declared(config) |
| 312 | ): |
| 313 | # Host-fetched X lane: the model passes connector results through |
| 314 | # --x-posts, so X is served without an engine backend. |
| 315 | available.append("x") |
| 316 | else: |
| 317 | # Safe inspection (--diagnose/--preflight) skips browser-cookie |
| 318 | # extraction, so get_x_source is None even though a real run would |
| 319 | # authenticate X via FROM_BROWSER. Report it as available so consumers |
| 320 | # of available_sources (SKILL.md ACTIVE_SOURCES_LIST) don't under-report. |
| 321 | # diagnose() precomputes the predicate and passes it via x_pending to |
| 322 | # avoid evaluating it twice in one diagnose() call. |
| 323 | if x_pending is None: |
| 324 | x_pending = env.x_pending_browser_auth(config) |
| 325 | if x_pending: |
| 326 | available.append("x") |
| 327 | if which("yt-dlp") or env.is_youtube_sc_available(config): |
| 328 | available.append("youtube") |
| 329 | available.extend(["hackernews", "polymarket"]) |
| 330 | # StockTwits is gated to ticker/crypto topics only (flag set in run()). |
| 331 | if config.get("_financial_topic"): |
| 332 | available.append("stocktwits") |
| 333 | # GitHub is reachable via the unauthenticated REST tier too, so it is |
| 334 | # available even without a token/gh CLI (a token only raises rate limits). |
| 335 | available.append("github") |
| 336 | # DripStack is opt-in only (owner decision, #791): a commercial |
| 337 | # third-party API must never receive default-run traffic. Opt in per run |
| 338 | # (--search dripstack) or persistently (INCLUDE_SOURCES=dripstack in |
| 339 | # .env, the LinkedIn/Perplexity pattern); the search API is free and |
| 340 | # public (no key), so the opt-in itself is the gate. |
| 341 | include_sources = { |
| 342 | token.strip() |
| 343 | for token in (config.get("INCLUDE_SOURCES") or "").lower().split(",") |
| 344 | if token.strip() |
| 345 | } |
| 346 | if "dripstack" in include_sources or ( |
| 347 | requested_sources and "dripstack" in requested_sources |
| 348 | ): |
| 349 | available.append("dripstack") |
| 350 | if which("digg-pp-cli"): |
| 351 | available.append("digg") |
| 352 | # arXiv is default-on when its Printing Press CLI is installed (zero auth). |
| 353 | # The adapter relevance-and-recency gates so it stays quiet off-topic. |
| 354 | if which("arxiv-pp-cli"): |
| 355 | available.append("arxiv") |
| 356 | # Techmeme is default-on when its CLI is installed (zero auth; sub-second |
| 357 | # local sync before each run's first search). |
| 358 | if which("techmeme-pp-cli"): |
| 359 | available.append("techmeme") |
| 360 | if env.is_bluesky_available(config): |
| 361 | available.append("bluesky") |
| 362 | if env.is_truthsocial_available(config): |
| 363 | available.append("truthsocial") |
| 364 | # Grounding (general web) is available when a paid backend is configured OR |
| 365 | # the keyless floor is permitted (i.e. the host has no native search). On a |
| 366 | # native-search host with no paid key, keyless_web_allowed is False and the |
| 367 | # engine leaves general web to the model's own search. |
| 368 | if (config.get("BRAVE_API_KEY") or config.get("EXA_API_KEY") |
| 369 | or config.get("SERPER_API_KEY") or config.get("PARALLEL_API_KEY") |
| 370 | or env.keyless_web_allowed(config)): |
| 371 | available.append("grounding") |
| 372 | if requested_sources and "jobs" in requested_sources: |
| 373 | available.append("jobs") |
| 374 | # Perplexity Agent API: opt-in additive source via INCLUDE_SOURCES=perplexity |
| 375 | if _has_perplexity_provider(config) and ( |
| 376 | "perplexity" in include_sources or (requested_sources and "perplexity" in requested_sources) |
| 377 | ): |
| 378 | available.append("perplexity") |
| 379 | # LinkedIn: opt-in additive source via INCLUDE_SOURCES=linkedin (same |
| 380 | # consent pattern as Perplexity). Unlike tiktok/instagram, which are |
| 381 | # offered during SKILL.md Step 0 onboarding, LinkedIn is power-user-only |
| 382 | # and must not silently activate for existing SCRAPECREATORS_API_KEY |
| 383 | # holders. |
| 384 | if config.get("SCRAPECREATORS_API_KEY") and ( |
| 385 | "linkedin" in include_sources or (requested_sources and "linkedin" in requested_sources) |
| 386 | ): |
| 387 | available.append("linkedin") |
| 388 | # Trustpilot: opt-in additive source via INCLUDE_SOURCES=trustpilot (same |
| 389 | # consent pattern as Perplexity/LinkedIn). Off by default -- unlike arXiv and |
| 390 | # Techmeme, which are zero-auth, it can spawn a one-time headless-Chrome WAF |
| 391 | # cookie harvest on a brand topic, so activating it is the user's choice. |
| 392 | if which("trustpilot-pp-cli") and ( |
| 393 | "trustpilot" in include_sources or (requested_sources and "trustpilot" in requested_sources) |
| 394 | ): |
| 395 | available.append("trustpilot") |
| 396 | # Amazon: opt-in additive source, dual-gated. The Bright Data CLI must be |
| 397 | # on the agent subprocess PATH and carry a credential signal, AND the run |
| 398 | # must ask for it -- the model per-run via --search, or the user durably |
| 399 | # via INCLUDE_SOURCES=amazon. Never inferred from topic shape: the engine |
| 400 | # misroutes most shopping phrasings, and auto-firing would spend a CLI |
| 401 | # owner's credits on runs that have nothing to do with products. |
| 402 | if brightdata.is_available(config) and ( |
| 403 | "amazon" in include_sources or (requested_sources and "amazon" in requested_sources) |
| 404 | ): |
| 405 | available.append("amazon") |
| 406 | # Meta Ads: opt-in additive source on the Amazon precedent. The |
| 407 | # ScrapeCreators key must be present AND the run must ask for it -- the |
| 408 | # model per-run via --search, or the user durably via |
| 409 | # INCLUDE_SOURCES=meta_ads. Never inferred from topic shape: keyword ad |
| 410 | # search on a non-brand topic returns a wrong-entity advertiser, and |
| 411 | # auto-firing would spend credits resolving it. |
| 412 | if config.get("SCRAPECREATORS_API_KEY") and ( |
| 413 | "meta_ads" in include_sources |
| 414 | or (requested_sources and "meta_ads" in requested_sources) |
| 415 | ): |
| 416 | available.append("meta_ads") |
| 417 | if ( |
| 418 | "xiaohongshu" in include_sources |
| 419 | or (requested_sources and "xiaohongshu" in requested_sources) |
| 420 | ) and env.is_xiaohongshu_available(config): |
| 421 | available.append("xiaohongshu") |
| 422 | # Threads: opt-in via INCLUDE_SOURCES (same pattern as perplexity/linkedin). |
| 423 | # Was auto-on with the key; gated so the onboarding "Everything" tier is a |
| 424 | # real choice vs the "Recommended" (TikTok/Instagram) tier. |
| 425 | if env.is_threads_available(config) and ( |
| 426 | "threads" in include_sources or (requested_sources and "threads" in requested_sources) |
| 427 | ): |
| 428 | available.append("threads") |
| 429 | # Pinterest: opt-in via INCLUDE_SOURCES. Previously read requested_sources |
| 430 | # only, so a persisted INCLUDE_SOURCES=pinterest never activated it; now it |
| 431 | # honors both the per-run --sources list and the saved config. |
| 432 | if env.is_pinterest_available(config) and ( |
| 433 | "pinterest" in include_sources or (requested_sources and "pinterest" in requested_sources) |
| 434 | ): |
| 435 | available.append("pinterest") |
| 436 | # Telegram: opt-in via INCLUDE_SOURCES AND requires a channel list. The |
| 437 | # channel list (TELEGRAM_SOURCES env or --telegram-sources CLI) is the gate: |
| 438 | # without named channels there is no discovery endpoint to call. |
| 439 | if config.get("SCRAPECREATORS_API_KEY") and ( |
| 440 | "telegram" in include_sources or (requested_sources and "telegram" in requested_sources) |
| 441 | ): |
| 442 | if telegram.is_telegram_configured(config): |
| 443 | available.append("telegram") |
| 444 | # xquik is a backend of the single "x" source (see env.x_backend_chain), |
| 445 | # not a separate parallel source — registered via the "x" entry above. |
| 446 | exclude = {s.strip().lower() for s in (config.get("EXCLUDE_SOURCES") or "").split(",") if s.strip()} |
| 447 | if exclude: |
| 448 | available = [s for s in available if s not in exclude] |
| 449 | return available |
| 450 | |
| 451 | |
| 452 | def _mock_discovery_items( |
| 453 | source: str, |
| 454 | domain: str, |
| 455 | to_date: str, |
| 456 | ) -> list[dict[str, Any]]: |
| 457 | """Deterministic listing fixtures for the public --mock CLI contract.""" |
| 458 | labels = [ |
| 459 | "Agent memory protocols", |
| 460 | "Browser-using agents", |
| 461 | "Local agent runtimes", |
| 462 | "Multi-agent orchestration", |
| 463 | "Agent security sandboxes", |
| 464 | "Voice agent latency", |
| 465 | ] |
| 466 | end = datetime.fromisoformat(to_date).date() |
| 467 | items: list[dict[str, Any]] = [] |
| 468 | for index, label in enumerate(labels, start=1): |
| 469 | published = (end - timedelta(days=index)).isoformat() |
| 470 | slug = re.sub(r"[^a-z0-9]+", "-", label.lower()).strip("-") |
| 471 | if source == "reddit": |
| 472 | items.append({ |
| 473 | "id": f"discovery-r-{index}", |
| 474 | "title": label, |
| 475 | "url": f"https://reddit.com/r/example/comments/{slug}", |
| 476 | "subreddit": "example", |
| 477 | "date": published, |
| 478 | "engagement": {"score": 180 - index * 10, "num_comments": 30 + index}, |
| 479 | "selftext": label, |
| 480 | "relevance": 0.9, |
| 481 | "why_relevant": "Mock discovery listing", |
| 482 | }) |
| 483 | elif source == "hackernews": |
| 484 | items.append({ |
| 485 | "id": f"discovery-hn-{index}", |
| 486 | "title": label, |
| 487 | "url": f"https://example.com/{slug}", |
| 488 | "hn_url": f"https://news.ycombinator.com/item?id={index}", |
| 489 | "author": f"example{index}", |
| 490 | "date": published, |
| 491 | "engagement": {"points": 120 - index * 8, "comments": 20 + index}, |
| 492 | "relevance": 0.88, |
| 493 | "why_relevant": "Mock HN discovery listing", |
| 494 | }) |
| 495 | elif source == "digg": |
| 496 | items.append({ |
| 497 | "id": f"discovery-d-{index}", |
| 498 | "title": label, |
| 499 | "url": f"https://di.gg/ai/{slug}", |
| 500 | "tldr": label, |
| 501 | "date": published, |
| 502 | "engagement": {"postCount": 30 - index, "uniqueAuthors": 12 - index}, |
| 503 | "relevance": 0.9, |
| 504 | "why_relevant": "Mock Digg discovery cluster", |
| 505 | }) |
| 506 | elif source == "x": |
| 507 | items.append({ |
| 508 | "id": f"discovery-x-{index}", |
| 509 | "text": label, |
| 510 | "url": f"https://x.com/example{index}/status/{index}", |
| 511 | "author_handle": f"example{index}", |
| 512 | "date": published, |
| 513 | "engagement": {"likes": 140 - index * 9, "reposts": 18 + index}, |
| 514 | "relevance": 0.9, |
| 515 | "why_relevant": "Mock X discovery activity", |
| 516 | }) |
| 517 | return items |
| 518 | |
| 519 | |
| 520 | def _matches_discovery_domain(domain: str, text: str) -> bool: |
| 521 | """Require a distinctive domain term, not a generic token such as ``AI``.""" |
| 522 | def terms(value: str) -> set[str]: |
| 523 | # Keep BOTH the surface form and the naive stem: replacing the token |
| 524 | # broke non-plurals ("bias" -> "bia", "crisis" -> "crisi") so in-domain |
| 525 | # listings stopped intersecting. The union preserves plural matching |
| 526 | # without corrupting the anchor. |
| 527 | words: set[str] = set() |
| 528 | for word in relevance.tokenize(value): |
| 529 | words.add(word) |
| 530 | if len(word) > 4 and word.endswith("s") and not word.endswith("ss"): |
| 531 | words.add(word[:-1]) |
| 532 | return words |
| 533 | |
| 534 | domain_terms = terms(domain) |
| 535 | anchors = domain_terms - _DISCOVERY_GENERIC_DOMAIN_TERMS |
| 536 | return bool((anchors or domain_terms) & terms(text)) |
| 537 | |
| 538 | |
| 539 | def _fetch_discovery_source( |
| 540 | source: str, |
| 541 | plan: schema.DiscoveryPlan, |
| 542 | *, |
| 543 | from_date: str, |
| 544 | to_date: str, |
| 545 | depth: str, |
| 546 | mock: bool, |
| 547 | config: dict[str, Any], |
| 548 | keyword_gate: bool = True, |
| 549 | ) -> tuple[list[dict[str, Any]], str | None]: |
| 550 | """Fetch one listing/river source for the nominate stage. |
| 551 | |
| 552 | ``keyword_gate`` controls whether items are filtered to the domain by |
| 553 | ``_matches_discovery_domain``. Domain-scoped discovery (``--discover X``) |
| 554 | keeps the gate on; global trending (``--discover`` with no domain) turns it |
| 555 | off, because there is no keyword to gate against - the river feeds ARE the |
| 556 | "what is hot right now" signal, and the confidence floor downstream is what |
| 557 | keeps junk out, not a keyword match. |
| 558 | """ |
| 559 | if mock: |
| 560 | return _mock_discovery_items(source, plan.domain, to_date), None |
| 561 | if source == "reddit": |
| 562 | result = reddit_listing.fetch_discovery_listings( |
| 563 | plan.subreddits, depth=depth, query=plan.domain, |
| 564 | ) |
| 565 | items = result.get("items") or [] |
| 566 | if keyword_gate: |
| 567 | items = [ |
| 568 | item for item in items |
| 569 | if _matches_discovery_domain( |
| 570 | plan.domain, |
| 571 | f"{item.get('title') or ''} {item.get('selftext') or ''}", |
| 572 | ) |
| 573 | ] |
| 574 | return items, "; ".join(result.get("errors") or []) or None |
| 575 | if source == "hackernews": |
| 576 | result = hackernews.fetch_discovery_listings(from_date, to_date, depth=depth) |
| 577 | items = result.get("items") or [] |
| 578 | for item in items: |
| 579 | item["relevance"] = relevance.token_overlap_relevance( |
| 580 | plan.domain, |
| 581 | str(item.get("title") or ""), |
| 582 | ) |
| 583 | # HN is a broad technology listing, so keep only domain-bearing stories |
| 584 | # when a domain is in play; global trending keeps the whole front page. |
| 585 | if keyword_gate: |
| 586 | items = [ |
| 587 | item for item in items |
| 588 | if _matches_discovery_domain(plan.domain, str(item.get("title") or "")) |
| 589 | ] |
| 590 | errors = result.get("errors") or [] |
| 591 | return items, "; ".join(errors) or None |
| 592 | if source == "digg": |
| 593 | result = digg.search_digg(plan.domain, from_date, to_date, depth=depth) |
| 594 | items = digg.parse_digg_response(result, query=plan.domain) |
| 595 | # Digg is an AI-focused broad listing, so keep only domain-bearing |
| 596 | # clusters when scoped; global trending keeps the whole feed. |
| 597 | if keyword_gate: |
| 598 | items = [ |
| 599 | item for item in items |
| 600 | if _matches_discovery_domain(plan.domain, str(item.get("title") or "")) |
| 601 | ] |
| 602 | return items, result.get("error") |
| 603 | if source == "x": |
| 604 | # Discovery uses domain directly as query (no planner search_query) |
| 605 | query = plan.domain |
| 606 | last_error = "" |
| 607 | for backend in env.x_backend_chain(config): |
| 608 | items, error = _fetch_x_backend( |
| 609 | backend, query, from_date, to_date, depth, config, |
| 610 | ) |
| 611 | if items: |
| 612 | # Earlier failed-over backends' errors are observability, not |
| 613 | # degradation - but the producing backend's own error means |
| 614 | # these items are partial and must surface as such. |
| 615 | if last_error: |
| 616 | print(f"[x] earlier backend failed: {last_error}", file=sys.stderr) |
| 617 | return items, error or None |
| 618 | if error: |
| 619 | last_error = f"{backend}: {error}" |
| 620 | return [], last_error or None |
| 621 | raise ValueError(f"Unsupported discovery source: {source}") |
| 622 | |
| 623 | |
| 624 | def _discovery_engagement( |
| 625 | items: list[schema.SourceItem], |
| 626 | ) -> dict[str, dict[str, float | int]]: |
| 627 | totals: dict[str, dict[str, float | int]] = {} |
| 628 | for item in items: |
| 629 | bucket = totals.setdefault(item.source, {}) |
| 630 | for field, value in item.engagement.items(): |
| 631 | if not isinstance(value, (int, float)) or isinstance(value, bool): |
| 632 | continue |
| 633 | # Rank/score/reach metadata is not additive engagement: summing |
| 634 | # Digg ranks across items fabricates a metric (agent-export uses |
| 635 | # the same counter-field rule). |
| 636 | if not schema._is_counter_field(field): |
| 637 | continue |
| 638 | bucket[field] = bucket.get(field, 0) + value |
| 639 | return { |
| 640 | source: dict(sorted(metrics.items())) |
| 641 | for source, metrics in sorted(totals.items()) |
| 642 | } |
| 643 | |
| 644 | |
| 645 | def _discovery_momentum(items: list[schema.SourceItem], to_date: str) -> str: |
| 646 | as_of = datetime.fromisoformat(to_date).date() |
| 647 | ages: list[int] = [] |
| 648 | for item in items: |
| 649 | try: |
| 650 | published = datetime.fromisoformat((item.published_at or "").replace("Z", "+00:00")).date() |
| 651 | except (TypeError, ValueError): |
| 652 | continue |
| 653 | ages.append(max(0, (as_of - published).days)) |
| 654 | return "new-this-week" if ages and max(ages) < 7 else "building" |
| 655 | |
| 656 | |
| 657 | def nominate_candidates( |
| 658 | plan: schema.DiscoveryPlan, |
| 659 | *, |
| 660 | from_date: str, |
| 661 | to_date: str, |
| 662 | depth: str, |
| 663 | mock: bool, |
| 664 | config: dict[str, Any], |
| 665 | lookback_days: int, |
| 666 | keyword_gate: bool = True, |
| 667 | ) -> schema.RetrievalBundle: |
| 668 | """Stage 1 of discovery: fetch, normalize, and bundle candidate hot items |
| 669 | from the river/listing feeds. |
| 670 | |
| 671 | This is the topic-nomination pass. For domain discovery ``keyword_gate`` is |
| 672 | on and the feeds are filtered to the domain; for global trending it is off |
| 673 | and the feeds' own hot ranking IS the signal. The returned bundle feeds the |
| 674 | clustering + enrichment stages downstream. Every source's failure is |
| 675 | recorded on the bundle (never raised) so a single dead feed cannot sink the |
| 676 | run - the confidence floor decides whether the surviving evidence is enough. |
| 677 | """ |
| 678 | bundle = schema.RetrievalBundle() |
| 679 | with ThreadPoolExecutor(max_workers=max(1, len(plan.sources))) as executor: |
| 680 | futures = { |
| 681 | executor.submit( |
| 682 | _fetch_discovery_source, |
| 683 | source, |
| 684 | plan, |
| 685 | from_date=from_date, |
| 686 | to_date=to_date, |
| 687 | depth=depth, |
| 688 | mock=mock, |
| 689 | config=config, |
| 690 | keyword_gate=keyword_gate, |
| 691 | ): source |
| 692 | for source in plan.sources |
| 693 | } |
| 694 | for future in as_completed(futures): |
| 695 | source = futures[future] |
| 696 | bundle.mark_attempted(source) |
| 697 | try: |
| 698 | raw_items, partial_error = future.result() |
| 699 | normalized = normalize.normalize_source_items( |
| 700 | source, |
| 701 | raw_items, |
| 702 | from_date, |
| 703 | to_date, |
| 704 | freshness_mode="breaking", |
| 705 | ) |
| 706 | # Global trending has no domain; annotate against a neutral |
| 707 | # phrase so snippet extraction still works without biasing |
| 708 | # relevance toward any keyword. |
| 709 | prepared = relevance.PreparedQuery(plan.domain or "trending now") |
| 710 | normalized = signals.annotate_stream( |
| 711 | normalized, |
| 712 | prepared, |
| 713 | "breaking", |
| 714 | reference_date=to_date, |
| 715 | max_days=lookback_days, |
| 716 | ) |
| 717 | normalized = dedupe.dedupe_items(normalized) |
| 718 | for item in normalized: |
| 719 | item.snippet = snippet.extract_best_snippet(item, prepared) |
| 720 | bundle.add_items("discovery-listings", source, normalized) |
| 721 | if partial_error: |
| 722 | failure_state = ( |
| 723 | bird_x.classify_run_failure(partial_error) |
| 724 | if source == "x" and partial_error.startswith("bird:") |
| 725 | else http.classify_failure(message=partial_error) |
| 726 | ) |
| 727 | bundle.record_failure( |
| 728 | source, |
| 729 | failure_state, |
| 730 | partial_error, |
| 731 | ) |
| 732 | except Exception as exc: |
| 733 | state, attempted = _classify_source_failure(exc) |
| 734 | bundle.record_failure(source, state, str(exc), attempted=attempted) |
| 735 | return bundle |
| 736 | |
| 737 | |
| 738 | @dataclass(frozen=True) |
| 739 | class Nomination: |
| 740 | """A named candidate topic produced by the nominate stage. |
| 741 | |
| 742 | ``seed_score`` is the cheap pre-enrichment rank - seed velocity on the |
| 743 | nominate stage, blended with the HOST judge's content-worthiness on the |
| 744 | protocol resume leg (see ``rerank.judge_blended_score``). Enough to |
| 745 | decide WHICH candidates deserve a full pipeline pass, but not the final |
| 746 | ranking signal (that comes from enriched evidence downstream). |
| 747 | ``junk_shape`` flags help-me/beginner/musing shapes that should not |
| 748 | become content topics; ``worthiness`` is the host judge's 0-100 content |
| 749 | score, None on the heuristic path. |
| 750 | """ |
| 751 | |
| 752 | name: str |
| 753 | seed_score: float |
| 754 | items: list[schema.SourceItem] = field(default_factory=list) |
| 755 | summary: str = "" |
| 756 | junk_shape: bool = False |
| 757 | worthiness: float | None = None |
| 758 | |
| 759 | |
| 760 | def _cluster_entity_counts( |
| 761 | cluster: schema.Cluster, |
| 762 | candidate_map: dict[str, schema.Candidate], |
| 763 | ) -> Counter: |
| 764 | """Entity-token frequencies across a cluster's members (title + snippet).""" |
| 765 | counts: Counter = Counter() |
| 766 | for candidate_id in cluster.candidate_ids: |
| 767 | candidate = candidate_map.get(candidate_id) |
| 768 | if candidate: |
| 769 | counts.update(entity_extract.extract_text_entities( |
| 770 | f"{candidate.title} {candidate.snippet}" |
| 771 | )) |
| 772 | return counts |
| 773 | |
| 774 | |
| 775 | # Bound on how many distinguishing entity tokens a colliding cluster may try |
| 776 | # before it is treated as indistinguishable from the earlier story. Keeps a |
| 777 | # pathological cluster (dozens of unique tokens, every resulting name already |
| 778 | # taken) from scanning its whole vocabulary. |
| 779 | _DISAMBIGUATION_TOKEN_LIMIT = 5 |
| 780 | |
| 781 | |
| 782 | def _disambiguated_topic_name( |
| 783 | name: str, |
| 784 | cluster: schema.Cluster, |
| 785 | earlier_cluster: schema.Cluster, |
| 786 | candidate_map: dict[str, schema.Candidate], |
| 787 | entity_counts_cache: dict[str, Counter], |
| 788 | taken_names: dict[str, schema.Cluster], |
| 789 | ) -> str | None: |
| 790 | """Disambiguate a colliding topic name by appending the later cluster's |
| 791 | strongest entity token that the earlier cluster does not share. |
| 792 | |
| 793 | Distinguishing tokens are tried in descending strength order (bounded at |
| 794 | ``_DISAMBIGUATION_TOKEN_LIMIT``) and the first resulting name not already |
| 795 | present in ``taken_names`` (casefolded keys) wins: a first-choice suffix |
| 796 | colliding with an already-taken name must not drop a distinct story while |
| 797 | another distinguishing token remains. |
| 798 | |
| 799 | ``entity_counts_cache`` (keyed by cluster id, owned by the caller) memoizes |
| 800 | per-cluster entity counts so repeated collisions against the same cluster |
| 801 | never recompute them. |
| 802 | |
| 803 | Returns None when no distinguishing entity yields an unused name - the |
| 804 | clusters cannot be told apart by content, so the caller treats them as the |
| 805 | same story. |
| 806 | """ |
| 807 | def cached_counts(target: schema.Cluster) -> Counter: |
| 808 | counts = entity_counts_cache.get(target.cluster_id) |
| 809 | if counts is None: |
| 810 | counts = _cluster_entity_counts(target, candidate_map) |
| 811 | entity_counts_cache[target.cluster_id] = counts |
| 812 | return counts |
| 813 | |
| 814 | later_counts = cached_counts(cluster) |
| 815 | earlier_entities = set(cached_counts(earlier_cluster)) |
| 816 | name_tokens = {token.casefold() for token in name.split()} |
| 817 | choices = [ |
| 818 | (count, token) for token, count in later_counts.items() |
| 819 | if token not in earlier_entities and token.casefold() not in name_tokens |
| 820 | ] |
| 821 | # Strongest first = most frequent across the cluster; alphabetical |
| 822 | # tie-break keeps the result deterministic. |
| 823 | ranked = sorted(choices, key=lambda entry: (-entry[0], entry[1])) |
| 824 | for _, token in ranked[:_DISAMBIGUATION_TOKEN_LIMIT]: |
| 825 | display = token |
| 826 | for candidate_id in cluster.candidate_ids: |
| 827 | candidate = candidate_map.get(candidate_id) |
| 828 | if candidate is None: |
| 829 | continue |
| 830 | match = next( |
| 831 | ( |
| 832 | word.strip("\"'`()[]{}.,:;!?") |
| 833 | for word in f"{candidate.title} {candidate.snippet}".split() |
| 834 | if word.strip("\"'`()[]{}.,:;!?").lower() == token |
| 835 | ), |
| 836 | None, |
| 837 | ) |
| 838 | if match: |
| 839 | display = match |
| 840 | break |
| 841 | resolved = f"{name} {display}" |
| 842 | if resolved.casefold() not in taken_names: |
| 843 | return resolved |
| 844 | return None |
| 845 | |
| 846 | |
| 847 | def nominate_topic_pool( |
| 848 | bundle: schema.RetrievalBundle, |
| 849 | query_plan: schema.QueryPlan, |
| 850 | plan: schema.DiscoveryPlan, |
| 851 | *, |
| 852 | from_date: str, |
| 853 | to_date: str, |
| 854 | limit: int, |
| 855 | ) -> list[tuple[Nomination, str]]: |
| 856 | """Stage 1b of discovery: cluster nominated items into named candidate |
| 857 | topics, rank them, and pair each with its source cluster id. |
| 858 | |
| 859 | This is the shared core behind ``nominate_topics`` (the one-shot path, |
| 860 | which drops the cluster ids) and the leg-1 nominate-only sweep (which |
| 861 | keys nominations-bundle rows on them, see ``run_discover_nominate``). |
| 862 | |
| 863 | Naming and junk classification are the deterministic ``topic_shape`` |
| 864 | heuristics and ranking is velocity-only - the engine runs no LLM here. |
| 865 | Reasoning-model judgment lives in the host-judged protocol: the host |
| 866 | renames, junk-filters, and worthiness-scores this pool from the leg-1 |
| 867 | bundle, and ``run_discover_resume`` applies those verdicts. The one-shot |
| 868 | path ships the heuristic names as-is. |
| 869 | |
| 870 | Casefold name collisions are disambiguated (the later cluster's strongest |
| 871 | non-shared entity token is appended, trying successive tokens when the |
| 872 | first-choice suffix is itself already taken) rather than blindly dropped: |
| 873 | short distilled names collide far more often than raw 96-char titles, and |
| 874 | a silent drop hides a distinct story. A colliding cluster is dropped only |
| 875 | when it shares a representative candidate with the earlier one (the same |
| 876 | story surfacing twice) or when no distinguishing entity token yields an |
| 877 | unused name. |
| 878 | |
| 879 | Returns at most ``limit`` ``(nomination, cluster_id)`` pairs, never |
| 880 | padded - fewer clusters than ``limit`` means a shorter list, and the |
| 881 | confidence floor downstream decides whether what survived is worth |
| 882 | showing. |
| 883 | """ |
| 884 | candidates = weighted_rrf( |
| 885 | bundle.items_by_source_and_query, |
| 886 | query_plan, |
| 887 | pool_limit=80, |
| 888 | range_from=from_date, |
| 889 | range_to=to_date, |
| 890 | ) |
| 891 | for candidate in candidates: |
| 892 | velocity = rerank.discovery_velocity_score(candidate.source_items, as_of_date=to_date) |
| 893 | candidate.final_score = min(100.0, 12.0 * math.log1p(velocity)) if velocity else 0.0 |
| 894 | candidates.sort(key=lambda candidate: (-candidate.final_score, candidate.title.lower())) |
| 895 | clusters = cluster_candidates(candidates, query_plan) |
| 896 | candidate_map = {candidate.candidate_id: candidate for candidate in candidates} |
| 897 | |
| 898 | ranked_clusters: list[tuple[float, schema.Cluster, list[schema.SourceItem]]] = [] |
| 899 | for cluster in clusters: |
| 900 | cluster_items: list[schema.SourceItem] = [] |
| 901 | for candidate_id in cluster.candidate_ids: |
| 902 | candidate = candidate_map.get(candidate_id) |
| 903 | if candidate: |
| 904 | cluster_items.extend(candidate.source_items) |
| 905 | score = rerank.discovery_velocity_score(cluster_items, as_of_date=to_date) |
| 906 | if score <= 0: |
| 907 | continue |
| 908 | ranked_clusters.append((score, cluster, cluster_items)) |
| 909 | ranked_clusters.sort(key=lambda entry: (-entry[0], entry[1].title.lower())) |
| 910 | |
| 911 | # Heuristic naming from each cluster's leader text (title + snippet). |
| 912 | named: list[tuple[float, schema.Cluster, list[schema.SourceItem], str, bool]] = [] |
| 913 | for score, cluster, cluster_items in ranked_clusters: |
| 914 | leader = candidate_map.get(cluster.representative_ids[0]) if cluster.representative_ids else None |
| 915 | title = (leader.title if leader else cluster.title) or "" |
| 916 | snip = (leader.snippet if leader else "") or "" |
| 917 | name = topic_shape.distill_topic_name(title, snip) or plan.domain or title |
| 918 | junk_shape = topic_shape.is_junk_shape(title, snip) |
| 919 | named.append((score, cluster, cluster_items, name, junk_shape)) |
| 920 | named.sort(key=lambda entry: (-entry[0], entry[3].lower())) |
| 921 | |
| 922 | pool: list[tuple[Nomination, str]] = [] |
| 923 | taken_names: dict[str, schema.Cluster] = {} |
| 924 | entity_counts_cache: dict[str, Counter] = {} |
| 925 | for score, cluster, cluster_items, name, junk_shape in named: |
| 926 | name_key = name.casefold() |
| 927 | if name_key in taken_names: |
| 928 | earlier_cluster = taken_names[name_key] |
| 929 | if set(cluster.representative_ids) & set(earlier_cluster.representative_ids): |
| 930 | continue # same story surfacing twice |
| 931 | resolved = _disambiguated_topic_name( |
| 932 | name, cluster, earlier_cluster, candidate_map, entity_counts_cache, |
| 933 | taken_names, |
| 934 | ) |
| 935 | if resolved is None: |
| 936 | continue # indistinguishable by content: treat as the same story |
| 937 | name = resolved |
| 938 | name_key = name.casefold() |
| 939 | taken_names[name_key] = cluster |
| 940 | leader = candidate_map.get(cluster.representative_ids[0]) if cluster.representative_ids else None |
| 941 | summary = (leader.snippet if leader else "") or (leader.title if leader else name) |
| 942 | pool.append((Nomination( |
| 943 | name=name, |
| 944 | seed_score=score, |
| 945 | items=cluster_items, |
| 946 | summary=summary, |
| 947 | junk_shape=junk_shape, |
| 948 | ), cluster.cluster_id)) |
| 949 | if len(pool) >= limit: |
| 950 | break |
| 951 | return pool |
| 952 | |
| 953 | |
| 954 | def nominate_topics( |
| 955 | bundle: schema.RetrievalBundle, |
| 956 | query_plan: schema.QueryPlan, |
| 957 | plan: schema.DiscoveryPlan, |
| 958 | *, |
| 959 | from_date: str, |
| 960 | to_date: str, |
| 961 | limit: int, |
| 962 | ) -> list[Nomination]: |
| 963 | """``nominate_topic_pool`` without the cluster ids: the one-shot |
| 964 | discovery path's contract (see that function for the full semantics).""" |
| 965 | return [ |
| 966 | nomination |
| 967 | for nomination, _cluster_id in nominate_topic_pool( |
| 968 | bundle, query_plan, plan, from_date=from_date, to_date=to_date, limit=limit, |
| 969 | ) |
| 970 | ] |
| 971 | |
| 972 | |
| 973 | # Enrichment fan-out bounds. Sub-runs hit the same upstream APIs as a normal |
| 974 | # research pass, so parallelism stays low and the whole batch runs against a |
| 975 | # wall-clock budget - a slow topic is dropped, never fatal. |
| 976 | ENRICH_LIMIT = 6 |
| 977 | ENRICH_DEPTH = "quick" |
| 978 | ENRICH_MAX_WORKERS = 3 |
| 979 | ENRICH_BUDGET_SECONDS = 240.0 |
| 980 | |
| 981 | |
| 982 | @dataclass |
| 983 | class EnrichedTopic: |
| 984 | """A nomination plus the full-pipeline evidence gathered for it. |
| 985 | |
| 986 | ``report`` is None when enrichment for this topic failed or ran past the |
| 987 | batch budget - the topic survives as nomination-only and the confidence |
| 988 | floor downstream decides whether its seed evidence is enough to show. |
| 989 | """ |
| 990 | |
| 991 | nomination: Nomination |
| 992 | report: schema.Report | None = None |
| 993 | error: str | None = None |
| 994 | |
| 995 | |
| 996 | def enrich_nominations( |
| 997 | nominations: list[Nomination], |
| 998 | *, |
| 999 | config: dict[str, Any], |
| 1000 | requested_sources: list[str] | None = None, |
| 1001 | mock: bool = False, |
| 1002 | depth: str = ENRICH_DEPTH, |
| 1003 | lookback_days: int = 30, |
| 1004 | as_of_date: str | None = None, |
| 1005 | max_workers: int = ENRICH_MAX_WORKERS, |
| 1006 | budget_seconds: float = ENRICH_BUDGET_SECONDS, |
| 1007 | ) -> list[EnrichedTopic]: |
| 1008 | """Stage 2 of discovery: run the real research pipeline on each nomination. |
| 1009 | |
| 1010 | Each nominated topic gets a full ``run()`` pass (``internal_subrun=True``, |
| 1011 | same lane as comparison-mode sub-runs), which buys the whole multi-source |
| 1012 | corpus - Reddit with comments, X, YouTube, Techmeme, arXiv, HN, Polymarket, |
| 1013 | web - plus clustering and ranking, with zero bespoke fetch code. |
| 1014 | |
| 1015 | Failure containment: a topic whose sub-run raises is returned with |
| 1016 | ``report=None`` and the error recorded; topics still unfinished when the |
| 1017 | batch budget expires are likewise dropped to nomination-only. The batch |
| 1018 | never raises and preserves nomination order. |
| 1019 | """ |
| 1020 | if not nominations: |
| 1021 | return [] |
| 1022 | |
| 1023 | def _run_one(nomination: Nomination) -> schema.Report: |
| 1024 | # Per-worker copy: run() mutates config in place |
| 1025 | # (config["_financial_topic"] = ...), so sharing one dict across |
| 1026 | # daemon threads lets topic A's flag overwrite topic B's mid-run, |
| 1027 | # with stragglers mutating past budget expiry. Same idiom as the |
| 1028 | # competitor runner (entity_config = dict(config)). |
| 1029 | return run( |
| 1030 | topic=nomination.name, |
| 1031 | config=dict(config), |
| 1032 | depth=depth, |
| 1033 | requested_sources=requested_sources, |
| 1034 | mock=mock, |
| 1035 | lookback_days=lookback_days, |
| 1036 | as_of_date=as_of_date, |
| 1037 | internal_subrun=True, |
| 1038 | # Enrichment passes never carry a connector envelope, so the |
| 1039 | # per-session lane signal must not plan X in and record a |
| 1040 | # spurious X error on every nominated topic. |
| 1041 | suppress_x_host_lane=True, |
| 1042 | ) |
| 1043 | |
| 1044 | # Daemon threads + a semaphore instead of ThreadPoolExecutor: executor |
| 1045 | # threads are non-daemon and joined at interpreter shutdown, so one hung |
| 1046 | # sub-run could keep the whole process alive long after its topic was |
| 1047 | # downgraded to nomination-only. Daemon workers make the wall-clock budget |
| 1048 | # real - stragglers cannot delay process exit. Abandonment is safe because |
| 1049 | # internal_subrun passes write nothing to disk (no save, no library sync, |
| 1050 | # no store), and every fetch layer inside run() carries its own timeout. |
| 1051 | youtube_yt.reset_search_cache() |
| 1052 | enriched: dict[str, EnrichedTopic] = {} |
| 1053 | results_queue: queue.Queue[tuple[Nomination, schema.Report | None, Exception | None]] = queue.Queue() |
| 1054 | slots = threading.Semaphore(max(1, max_workers)) |
| 1055 | |
| 1056 | def _worker(nomination: Nomination) -> None: |
| 1057 | with slots: |
| 1058 | try: |
| 1059 | results_queue.put((nomination, _run_one(nomination), None)) |
| 1060 | except Exception as exc: # noqa: BLE001 - containment is the contract |
| 1061 | results_queue.put((nomination, None, exc)) |
| 1062 | |
| 1063 | for nomination in nominations: |
| 1064 | threading.Thread( |
| 1065 | target=_worker, |
| 1066 | args=(nomination,), |
| 1067 | name=f"discover-enrich-{nomination.name[:32]}", |
| 1068 | daemon=True, |
| 1069 | ).start() |
| 1070 | |
| 1071 | deadline = time.monotonic() + max(1.0, budget_seconds) |
| 1072 | pending = len(nominations) |
| 1073 | while pending and (remaining := deadline - time.monotonic()) > 0: |
| 1074 | try: |
| 1075 | nomination, report, exc = results_queue.get(timeout=min(remaining, 0.5)) |
| 1076 | except queue.Empty: |
| 1077 | continue |
| 1078 | pending -= 1 |
| 1079 | if exc is None: |
| 1080 | enriched[nomination.name] = EnrichedTopic( |
| 1081 | nomination=nomination, report=report, |
| 1082 | ) |
| 1083 | else: |
| 1084 | enriched[nomination.name] = EnrichedTopic( |
| 1085 | nomination=nomination, |
| 1086 | error=f"{type(exc).__name__}: {exc}", |
| 1087 | ) |
| 1088 | print( |
| 1089 | f"[Discover] enrichment failed for {nomination.name!r}: " |
| 1090 | f"{type(exc).__name__}: {exc}", |
| 1091 | file=sys.stderr, |
| 1092 | ) |
| 1093 | # Budget expired (or all done): unfinished topics fall through below as |
| 1094 | # nomination-only; their daemon workers are abandoned and cannot block exit. |
| 1095 | |
| 1096 | results: list[EnrichedTopic] = [] |
| 1097 | for nomination in nominations: |
| 1098 | entry = enriched.get(nomination.name) |
| 1099 | if entry is None: |
| 1100 | entry = EnrichedTopic( |
| 1101 | nomination=nomination, |
| 1102 | error="enrichment budget exhausted", |
| 1103 | ) |
| 1104 | print( |
| 1105 | f"[Discover] enrichment budget exhausted before {nomination.name!r} " |
| 1106 | "finished; keeping nomination-only evidence", |
| 1107 | file=sys.stderr, |
| 1108 | ) |
| 1109 | results.append(entry) |
| 1110 | return results |
| 1111 | |
| 1112 | |
| 1113 | def _enriched_evidence_items(entry: EnrichedTopic) -> list[schema.SourceItem]: |
| 1114 | """The items a topic is judged on: the enriched corpus when the pipeline |
| 1115 | pass succeeded, the nomination's seed items otherwise.""" |
| 1116 | if entry.report is not None: |
| 1117 | flattened: list[schema.SourceItem] = [] |
| 1118 | for source_items in entry.report.items_by_source.values(): |
| 1119 | flattened.extend(source_items) |
| 1120 | if flattened: |
| 1121 | return flattened |
| 1122 | return entry.nomination.items |
| 1123 | |
| 1124 | |
| 1125 | def _best_community_comment(items: list[schema.SourceItem]) -> str | None: |
| 1126 | """The strongest verbatim community comment across a topic's evidence, |
| 1127 | formatted with attribution - the voice-of-the-people line on a trend card. |
| 1128 | |
| 1129 | Vote strength is per-platform-normalized (signals.normalized_comment_vote) |
| 1130 | so one viral platform's counts don't drown out the rest. |
| 1131 | """ |
| 1132 | best: tuple[float, str, str | None, float | int | None] | None = None |
| 1133 | for item in items: |
| 1134 | comments = item.metadata.get("top_comments") or [] |
| 1135 | for comment in comments: |
| 1136 | if not isinstance(comment, dict): |
| 1137 | continue |
| 1138 | body = (comment.get("excerpt") or comment.get("text") or comment.get("body") or "").strip() |
| 1139 | if len(body) < 12: |
| 1140 | continue |
| 1141 | strength = signals.normalized_comment_vote(item.source, comment.get("score")) |
| 1142 | if best is None or strength > best[0]: |
| 1143 | best = (strength, body, comment.get("author"), comment.get("score")) |
| 1144 | if best is None: |
| 1145 | return None |
| 1146 | _, body, author, score = best |
| 1147 | # Comment bodies that themselves start/end with quote characters would |
| 1148 | # render as doubled quotes inside our wrapping quotes. |
| 1149 | body = body.strip('"“”‘’\'').strip() |
| 1150 | if len(body) > 200: |
| 1151 | body = body[:197].rsplit(" ", 1)[0] + "..." |
| 1152 | attribution = f" - {author}" if author else "" |
| 1153 | votes = ( |
| 1154 | f" ({int(score):,} votes)" |
| 1155 | if isinstance(score, (int, float)) and not isinstance(score, bool) and score > 0 |
| 1156 | else "" |
| 1157 | ) |
| 1158 | return f'"{body}"{attribution}{votes}' |
| 1159 | |
| 1160 | |
| 1161 | @dataclass(frozen=True) |
| 1162 | class _DiscoverySweep: |
| 1163 | """The shared front half of both discovery entry points: the resolved |
| 1164 | plan and window, the swept listing bundle, and finalized per-source |
| 1165 | status. Everything downstream (judging, enrichment, floor, queue) |
| 1166 | belongs to the caller's leg.""" |
| 1167 | |
| 1168 | plan: schema.DiscoveryPlan |
| 1169 | query_plan: schema.QueryPlan |
| 1170 | from_date: str |
| 1171 | to_date: str |
| 1172 | bundle: schema.RetrievalBundle |
| 1173 | source_status: dict[str, schema.SourceOutcome] |
| 1174 | |
| 1175 | |
| 1176 | def _discovery_sweep( |
| 1177 | *, |
| 1178 | domain: str, |
| 1179 | config: dict[str, Any], |
| 1180 | depth: str, |
| 1181 | requested_sources: list[str] | None, |
| 1182 | mock: bool, |
| 1183 | subreddits: list[str] | None, |
| 1184 | lookback_days: int, |
| 1185 | as_of_date: str | None, |
| 1186 | ) -> _DiscoverySweep: |
| 1187 | """Resolve the momentum window, validate/bound the listing sources, build |
| 1188 | the discovery plan, sweep the river feeds, and finalize source status. |
| 1189 | |
| 1190 | Shared verbatim by ``run_discover`` (one-shot) and |
| 1191 | ``run_discover_nominate`` (protocol leg 1) so the two paths can never |
| 1192 | drift on what a sweep means.""" |
| 1193 | from_date, to_date = dates.get_date_range(lookback_days, as_of_date=as_of_date) |
| 1194 | requested = normalize_requested_sources(requested_sources) |
| 1195 | unsupported = sorted(set(requested or []) - set(DISCOVERY_SOURCES)) |
| 1196 | if unsupported: |
| 1197 | raise ValueError( |
| 1198 | "Discovery supports listing sources only: reddit, hackernews, digg " |
| 1199 | f"(unsupported: {', '.join(unsupported)})" |
| 1200 | ) |
| 1201 | available = list(DISCOVERY_SOURCES) if mock else [ |
| 1202 | source for source in available_sources(config, requested, x_pending=False) |
| 1203 | if source in DISCOVERY_SOURCES |
| 1204 | ] |
| 1205 | if requested: |
| 1206 | available = [source for source in available if source in requested] |
| 1207 | plan = planner.build_discovery_plan( |
| 1208 | domain, |
| 1209 | available_sources=available, |
| 1210 | subreddits=subreddits, |
| 1211 | ) |
| 1212 | |
| 1213 | global_mode = not plan.domain |
| 1214 | domain_label = plan.domain or "everything" |
| 1215 | query_plan = schema.QueryPlan( |
| 1216 | intent="breaking_news", |
| 1217 | freshness_mode="breaking", |
| 1218 | cluster_mode="story", |
| 1219 | raw_topic=plan.domain, |
| 1220 | subqueries=[schema.SubQuery( |
| 1221 | label="discovery-listings", |
| 1222 | search_query=plan.domain, |
| 1223 | ranking_query=f"What is accelerating in {domain_label}?", |
| 1224 | sources=list(plan.sources), |
| 1225 | )], |
| 1226 | source_weights={source: 1.0 for source in plan.sources}, |
| 1227 | notes=["discover-mode", "listing-sweep"], |
| 1228 | ) |
| 1229 | |
| 1230 | bundle = nominate_candidates( |
| 1231 | plan, |
| 1232 | from_date=from_date, |
| 1233 | to_date=to_date, |
| 1234 | depth=depth, |
| 1235 | mock=mock, |
| 1236 | config=config, |
| 1237 | lookback_days=lookback_days, |
| 1238 | # Global trending has no keyword to gate against - the river feeds' own |
| 1239 | # hot ranking is the signal and the confidence floor culls the junk. |
| 1240 | keyword_gate=not global_mode, |
| 1241 | ) |
| 1242 | |
| 1243 | source_status: dict[str, schema.SourceOutcome] = {} |
| 1244 | for source in DISCOVERY_SOURCES: |
| 1245 | if source in bundle.source_status: |
| 1246 | continue |
| 1247 | detail = ( |
| 1248 | "Source is not configured for discovery." |
| 1249 | ) |
| 1250 | source_status[source] = schema.SourceOutcome( |
| 1251 | source=source, |
| 1252 | state=schema.SKIPPED_UNCONFIGURED, |
| 1253 | attempted=False, |
| 1254 | detail=detail, |
| 1255 | fix_hint="doctor", |
| 1256 | ) |
| 1257 | source_status.update(_finalize_source_status(bundle.source_status, bundle.items_by_source)) |
| 1258 | return _DiscoverySweep( |
| 1259 | plan=plan, |
| 1260 | query_plan=query_plan, |
| 1261 | from_date=from_date, |
| 1262 | to_date=to_date, |
| 1263 | bundle=bundle, |
| 1264 | source_status=source_status, |
| 1265 | ) |
| 1266 | |
| 1267 | |
| 1268 | def _degraded_discovery_sources( |
| 1269 | source_status: dict[str, schema.SourceOutcome], |
| 1270 | ) -> list[str]: |
| 1271 | """Sources whose outcome is neither clean nor an expected skip.""" |
| 1272 | return [ |
| 1273 | source for source, outcome_state in source_status.items() |
| 1274 | if outcome_state.state not in {health.OK, schema.NO_RESULTS, schema.SKIPPED_UNCONFIGURED} |
| 1275 | ] |
| 1276 | |
| 1277 | |
| 1278 | @dataclass(frozen=True) |
| 1279 | class DiscoverNominateResult: |
| 1280 | """Leg 1 output of the host-judged discovery protocol: the ranked judge |
| 1281 | pool as ``(nomination, cluster_id)`` pairs plus the sweep context the CLI |
| 1282 | needs to write the nominations bundle - or to render the nothing-solid |
| 1283 | brief when the pool is empty.""" |
| 1284 | |
| 1285 | plan: schema.DiscoveryPlan |
| 1286 | from_date: str |
| 1287 | to_date: str |
| 1288 | source_status: dict[str, schema.SourceOutcome] |
| 1289 | pool: list[tuple[Nomination, str]] |
| 1290 | |
| 1291 | |
| 1292 | def run_discover_nominate( |
| 1293 | *, |
| 1294 | domain: str, |
| 1295 | config: dict[str, Any], |
| 1296 | depth: str = "default", |
| 1297 | requested_sources: list[str] | None = None, |
| 1298 | mock: bool = False, |
| 1299 | subreddits: list[str] | None = None, |
| 1300 | lookback_days: int = 30, |
| 1301 | as_of_date: str | None = None, |
| 1302 | ) -> DiscoverNominateResult: |
| 1303 | """Protocol leg 1: sweep the listings and build the FULL judge pool. |
| 1304 | |
| 1305 | Same sweep and clustering as ``run_discover``, but the pool is cut at |
| 1306 | ``rerank.JUDGE_POOL_LIMIT`` (not the enrichment limit). Like every |
| 1307 | discovery path it is deterministic-heuristic: no provider is ever |
| 1308 | resolved, so names and junk flags are the ``topic_shape`` baselines the |
| 1309 | host judges against. No enrichment, no confidence floor, no queue |
| 1310 | writes - those belong to legs 2 and 3. |
| 1311 | """ |
| 1312 | sweep = _discovery_sweep( |
| 1313 | domain=domain, |
| 1314 | config=config, |
| 1315 | depth=depth, |
| 1316 | requested_sources=requested_sources, |
| 1317 | mock=mock, |
| 1318 | subreddits=subreddits, |
| 1319 | lookback_days=lookback_days, |
| 1320 | as_of_date=as_of_date, |
| 1321 | ) |
| 1322 | pool = nominate_topic_pool( |
| 1323 | sweep.bundle, sweep.query_plan, sweep.plan, |
| 1324 | from_date=sweep.from_date, |
| 1325 | to_date=sweep.to_date, |
| 1326 | limit=rerank.JUDGE_POOL_LIMIT, |
| 1327 | ) |
| 1328 | return DiscoverNominateResult( |
| 1329 | plan=sweep.plan, |
| 1330 | from_date=sweep.from_date, |
| 1331 | to_date=sweep.to_date, |
| 1332 | source_status=sweep.source_status, |
| 1333 | pool=pool, |
| 1334 | ) |
| 1335 | |
| 1336 | |
| 1337 | def nominate_nothing_solid_report(result: DiscoverNominateResult) -> schema.DiscoveryReport: |
| 1338 | """The honest-empty leg-1 report: a zero-nomination sweep renders the |
| 1339 | same nothing-solid brief a one-shot run would (and writes no bundle).""" |
| 1340 | warnings = [ |
| 1341 | "The listing sweep nominated no topics this window; reporting " |
| 1342 | "nothing solid instead of ranked noise." |
| 1343 | ] |
| 1344 | failed = _degraded_discovery_sources(result.source_status) |
| 1345 | if failed: |
| 1346 | warnings.append(f"Some discovery sources degraded: {', '.join(sorted(failed))}.") |
| 1347 | return schema.DiscoveryReport( |
| 1348 | domain=result.plan.domain, |
| 1349 | range_from=result.from_date, |
| 1350 | range_to=result.to_date, |
| 1351 | generated_at=datetime.now(timezone.utc).isoformat(), |
| 1352 | plan=result.plan, |
| 1353 | topics=[], |
| 1354 | source_status=result.source_status, |
| 1355 | warnings=warnings, |
| 1356 | outcome="nothing-solid", |
| 1357 | weak_signal=None, |
| 1358 | ) |
| 1359 | |
| 1360 | |
| 1361 | def _floor_survivor_records( |
| 1362 | enriched_entries: list[EnrichedTopic], |
| 1363 | *, |
| 1364 | to_date: str, |
| 1365 | topic_limit: int, |
| 1366 | ) -> tuple[ |
| 1367 | list[dict[str, Any]], |
| 1368 | tuple[float, str] | None, |
| 1369 | tuple[float, str] | None, |
| 1370 | ]: |
| 1371 | """Apply the discovery confidence floor to enriched entries in order, |
| 1372 | returning the survivor records plus the strongest non-junk and junk weak |
| 1373 | signals among the failures. |
| 1374 | |
| 1375 | Shared verbatim by ``run_discover`` (one-shot) and ``run_discover_resume`` |
| 1376 | (protocol leg 2) so floor semantics can never drift between the paths. |
| 1377 | """ |
| 1378 | survivors: list[dict[str, Any]] = [] |
| 1379 | weak_signal: tuple[float, str] | None = None |
| 1380 | junk_weak_signal: tuple[float, str] | None = None |
| 1381 | for entry in enriched_entries: |
| 1382 | nomination = entry.nomination |
| 1383 | evidence_items = _enriched_evidence_items(entry) |
| 1384 | sources = sorted({item.source for item in evidence_items}) |
| 1385 | native_total = sum( |
| 1386 | rerank.discovery_engagement_total(item) for item in evidence_items |
| 1387 | ) |
| 1388 | score = rerank.discovery_velocity_score(evidence_items, as_of_date=to_date) |
| 1389 | if not rerank.passes_discovery_floor( |
| 1390 | source_count=len(sources), |
| 1391 | engagement_total=native_total, |
| 1392 | item_count=len(evidence_items), |
| 1393 | junk_shape=nomination.junk_shape, |
| 1394 | # Junk corroboration counts distinct SEED listing sources, never |
| 1395 | # the enriched corpus - a successful enrichment pass is |
| 1396 | # multi-source for almost any topic, so it would never bind. |
| 1397 | seed_source_count=len({item.source for item in nomination.items}), |
| 1398 | ): |
| 1399 | # Sub-floor evidence never ranks; remember what came closest so a |
| 1400 | # nothing-solid brief can still name the strongest weak signal. |
| 1401 | # Junk-shaped failures are tracked separately: the brief prefers |
| 1402 | # the strongest NON-junk failure and names a junk one only when |
| 1403 | # every failure is junk-shaped (never empty when failures exist). |
| 1404 | if nomination.junk_shape: |
| 1405 | if junk_weak_signal is None or score > junk_weak_signal[0]: |
| 1406 | junk_weak_signal = (score, nomination.name) |
| 1407 | elif weak_signal is None or score > weak_signal[0]: |
| 1408 | weak_signal = (score, nomination.name) |
| 1409 | continue |
| 1410 | if len(survivors) >= topic_limit: |
| 1411 | break |
| 1412 | source_phrase = ", ".join(sources[:-1]) + ( |
| 1413 | f" and {sources[-1]}" if len(sources) > 1 else (sources[0] if sources else "the listings") |
| 1414 | ) |
| 1415 | noun = "evidence item" if entry.report is not None else "listing item" |
| 1416 | why = ( |
| 1417 | f"{len(evidence_items)} {noun}{'s' if len(evidence_items) != 1 else ''} on " |
| 1418 | f"{source_phrase} generated {native_total:,.0f} native interactions. " |
| 1419 | f"{nomination.summary[:220]}" |
| 1420 | ) |
| 1421 | top_comment = _best_community_comment(evidence_items) if entry.report is not None else None |
| 1422 | # Stage-2 angle input: the survivor's strongest evidence, enriched |
| 1423 | # corpus when the pipeline pass succeeded, seed items otherwise |
| 1424 | # (evidence_items already resolves that). |
| 1425 | top_titles = [ |
| 1426 | item.title.strip() |
| 1427 | for item in sorted( |
| 1428 | evidence_items, |
| 1429 | key=rerank.discovery_engagement_total, |
| 1430 | reverse=True, |
| 1431 | ) |
| 1432 | if item.title and item.title.strip() |
| 1433 | ][:3] |
| 1434 | survivors.append({ |
| 1435 | "name": nomination.name, |
| 1436 | "why": why, |
| 1437 | "momentum": _discovery_momentum(evidence_items, to_date), |
| 1438 | "velocity_score": round(score, 2), |
| 1439 | "sources": sources, |
| 1440 | "engagement_by_source": _discovery_engagement(evidence_items), |
| 1441 | "evidence_urls": list(dict.fromkeys(item.url for item in evidence_items if item.url))[:5], |
| 1442 | "top_comment": top_comment, |
| 1443 | "titles": "; ".join(top_titles), |
| 1444 | "engagement_phrase": f"{native_total:,.0f} native interactions across {source_phrase}", |
| 1445 | }) |
| 1446 | return survivors, weak_signal, junk_weak_signal |
| 1447 | |
| 1448 | |
| 1449 | def _fold_same_story_records(survivors: list[dict[str, Any]]) -> list[dict[str, Any]]: |
| 1450 | """Same-story fold + velocity ordering over floor-survivor records. |
| 1451 | |
| 1452 | Floor survivors that share enriched evidence are the SAME story wearing |
| 1453 | two judged names (the real-run failure: two topics quoting the identical |
| 1454 | 1,635-vote comment). Duplicates = identical non-None top comment OR >= 2 |
| 1455 | shared evidence URLs; the lower-velocity twin is dropped, and a winning |
| 1456 | replacement re-scans the kept list to a fixpoint so chained overlap |
| 1457 | (A~C~B) still collapses to one survivor. Selection stays seed-ordered |
| 1458 | upstream; this only prunes, then sorts by displayed velocity (stable) so |
| 1459 | rank 1 is the highest velocity_score. |
| 1460 | """ |
| 1461 | def _same_story(a: dict[str, Any], b: dict[str, Any]) -> bool: |
| 1462 | if a["top_comment"] is not None and a["top_comment"] == b["top_comment"]: |
| 1463 | return True |
| 1464 | return len(set(a["evidence_urls"]) & set(b["evidence_urls"])) >= 2 |
| 1465 | |
| 1466 | folded: list[dict[str, Any]] = [] |
| 1467 | for record in survivors: |
| 1468 | # Fold to a fixpoint: when the incoming record REPLACES a kept one, |
| 1469 | # the replacement may share evidence with entries the dropped record |
| 1470 | # never matched (three-way chains: A kept, C shares the comment with |
| 1471 | # A and URLs with B). The winner re-scans the remaining kept entries |
| 1472 | # until nothing matches, so one story always yields one survivor. |
| 1473 | incoming: dict[str, Any] | None = record |
| 1474 | while incoming is not None: |
| 1475 | dup_index = next( |
| 1476 | (index for index, kept in enumerate(folded) if _same_story(incoming, kept)), |
| 1477 | None, |
| 1478 | ) |
| 1479 | if dup_index is None: |
| 1480 | folded.append(incoming) |
| 1481 | break |
| 1482 | kept = folded[dup_index] |
| 1483 | if incoming["velocity_score"] > kept["velocity_score"]: |
| 1484 | folded.pop(dup_index) |
| 1485 | dropped_name, kept_name = kept["name"], incoming["name"] |
| 1486 | else: |
| 1487 | dropped_name, kept_name = incoming["name"], kept["name"] |
| 1488 | incoming = None # dropped; the kept entry stays in place |
| 1489 | log.source_log( |
| 1490 | "Discover", |
| 1491 | f"folded duplicate story {dropped_name!r} into {kept_name!r} (shared evidence)", |
| 1492 | tty_only=False, |
| 1493 | ) |
| 1494 | |
| 1495 | folded.sort(key=lambda record: record["velocity_score"], reverse=True) |
| 1496 | return folded |
| 1497 | |
| 1498 | |
| 1499 | def _records_to_discovery_topics( |
| 1500 | folded: list[dict[str, Any]], |
| 1501 | ) -> list[schema.DiscoveryTopic]: |
| 1502 | """Folded survivor records to ranked topics (ranks = 1-based positions).""" |
| 1503 | return [ |
| 1504 | schema.DiscoveryTopic( |
| 1505 | rank=position, |
| 1506 | name=record["name"], |
| 1507 | why_spiking=record["why"], |
| 1508 | momentum=record["momentum"], |
| 1509 | velocity_score=record["velocity_score"], |
| 1510 | sources=record["sources"], |
| 1511 | engagement_by_source=record["engagement_by_source"], |
| 1512 | command=f'/last30days "{record["name"].replace(chr(34), chr(39))}"', |
| 1513 | evidence_urls=record["evidence_urls"], |
| 1514 | top_comment=record["top_comment"], |
| 1515 | corroboration_count=len(record["sources"]), |
| 1516 | ) |
| 1517 | for position, record in enumerate(folded, start=1) |
| 1518 | ] |
| 1519 | |
| 1520 | |
| 1521 | def _discovery_report_warnings( |
| 1522 | topics: list[schema.DiscoveryTopic], |
| 1523 | outcome: str, |
| 1524 | source_status: dict[str, schema.SourceOutcome], |
| 1525 | ) -> list[str]: |
| 1526 | """Coverage warnings shared by the one-shot and resume discovery paths. |
| 1527 | The resume leg never re-sweeps: it passes the bundle's RESTORED leg-1 |
| 1528 | sweep status, so a degraded feed from the sweep still reaches the leg-2 |
| 1529 | report exactly as the one-shot reports it.""" |
| 1530 | warnings: list[str] = [] |
| 1531 | if outcome == "nothing-solid": |
| 1532 | warnings.append( |
| 1533 | "No topic cleared the discovery confidence floor this window; " |
| 1534 | "reporting nothing solid instead of ranked noise." |
| 1535 | ) |
| 1536 | elif len(topics) < 5: |
| 1537 | warnings.append("Fewer than five topic clusters cleared the confidence floor this window.") |
| 1538 | if topics and all(len(topic.sources) == 1 for topic in topics): |
| 1539 | warnings.append("Discovery evidence is single-source; configure Digg for broader confirmation.") |
| 1540 | failed = _degraded_discovery_sources(source_status) |
| 1541 | if failed: |
| 1542 | warnings.append(f"Some discovery sources degraded: {', '.join(sorted(failed))}.") |
| 1543 | return warnings |
| 1544 | |
| 1545 | |
| 1546 | def run_discover( |
| 1547 | *, |
| 1548 | domain: str, |
| 1549 | config: dict[str, Any], |
| 1550 | depth: str = "default", |
| 1551 | requested_sources: list[str] | None = None, |
| 1552 | mock: bool = False, |
| 1553 | subreddits: list[str] | None = None, |
| 1554 | lookback_days: int = 30, |
| 1555 | as_of_date: str | None = None, |
| 1556 | limit: int = 10, |
| 1557 | enrich: bool = False, |
| 1558 | enrich_requested_sources: list[str] | None = None, |
| 1559 | ) -> schema.DiscoveryReport: |
| 1560 | """Sweep category listings and rank the topics gaining velocity. |
| 1561 | |
| 1562 | ``requested_sources`` bounds the listing sweep (discovery-capable feeds |
| 1563 | only). ``enrich_requested_sources`` bounds the per-topic research passes: |
| 1564 | None means every available source - which is what lets Techmeme, arXiv, |
| 1565 | YouTube, Polymarket, and community comments reach discovery despite having |
| 1566 | no river feed of their own. Pass the user's original --search list here so |
| 1567 | an explicit source boundary holds through enrichment too. |
| 1568 | """ |
| 1569 | sweep = _discovery_sweep( |
| 1570 | domain=domain, |
| 1571 | config=config, |
| 1572 | depth=depth, |
| 1573 | requested_sources=requested_sources, |
| 1574 | mock=mock, |
| 1575 | subreddits=subreddits, |
| 1576 | lookback_days=lookback_days, |
| 1577 | as_of_date=as_of_date, |
| 1578 | ) |
| 1579 | plan = sweep.plan |
| 1580 | from_date, to_date = sweep.from_date, sweep.to_date |
| 1581 | source_status = sweep.source_status |
| 1582 | |
| 1583 | # The engine never names or angles topics with an LLM: the one-shot path |
| 1584 | # is deterministic-heuristic by design, and reasoning-model judgment |
| 1585 | # lives in the host-judged SKILL.md protocol. Say so loudly once per live |
| 1586 | # run; --mock stays silent (a deliberate mock run is not a degraded run). |
| 1587 | if not mock: |
| 1588 | log.source_log( |
| 1589 | "Discover", |
| 1590 | "one-shot run: topic names use deterministic heuristics and no " |
| 1591 | "content angles are generated - a reasoning-model host running " |
| 1592 | "the SKILL.md discovery protocol gets host-judged names, junk " |
| 1593 | "filtering, and podcast/X angles", |
| 1594 | tty_only=False, |
| 1595 | ) |
| 1596 | |
| 1597 | topic_limit = max(5, min(10, limit)) |
| 1598 | nominations = nominate_topics( |
| 1599 | sweep.bundle, sweep.query_plan, plan, |
| 1600 | from_date=from_date, |
| 1601 | to_date=to_date, |
| 1602 | limit=ENRICH_LIMIT if enrich else topic_limit, |
| 1603 | ) |
| 1604 | |
| 1605 | if enrich and nominations: |
| 1606 | enriched_entries = enrich_nominations( |
| 1607 | nominations, |
| 1608 | config=config, |
| 1609 | requested_sources=enrich_requested_sources, |
| 1610 | mock=mock, |
| 1611 | lookback_days=lookback_days, |
| 1612 | as_of_date=as_of_date, |
| 1613 | ) |
| 1614 | else: |
| 1615 | enriched_entries = [ |
| 1616 | EnrichedTopic(nomination=nomination) for nomination in nominations |
| 1617 | ] |
| 1618 | |
| 1619 | survivors, weak_signal, junk_weak_signal = _floor_survivor_records( |
| 1620 | enriched_entries, to_date=to_date, topic_limit=topic_limit, |
| 1621 | ) |
| 1622 | folded = _fold_same_story_records(survivors) |
| 1623 | # One-shot topics ship without angles (podcast_angle / x_article_angle |
| 1624 | # stay None and the renderer omits those lines): content angles are a |
| 1625 | # host-judged protocol deliverable, written on the finalize leg. |
| 1626 | topics = _records_to_discovery_topics(folded) |
| 1627 | |
| 1628 | if weak_signal is None: |
| 1629 | weak_signal = junk_weak_signal |
| 1630 | |
| 1631 | outcome = "ok" if topics else "nothing-solid" |
| 1632 | |
| 1633 | return schema.DiscoveryReport( |
| 1634 | domain=plan.domain, |
| 1635 | range_from=from_date, |
| 1636 | range_to=to_date, |
| 1637 | generated_at=datetime.now(timezone.utc).isoformat(), |
| 1638 | plan=plan, |
| 1639 | topics=topics, |
| 1640 | source_status=source_status, |
| 1641 | warnings=_discovery_report_warnings(topics, outcome, source_status), |
| 1642 | outcome=outcome, |
| 1643 | weak_signal=weak_signal[1] if weak_signal and not topics else None, |
| 1644 | ) |
| 1645 | |
| 1646 | |
| 1647 | # Protocol leg 2 (resume) deep-tier enrichment bounds. The module-level |
| 1648 | # ENRICH_* constants above stay the one-shot --discover contract (quick depth, |
| 1649 | # 240s budget, 3 workers); a deep-tier bundle upgrades its per-topic sub-runs |
| 1650 | # to the default research depth with a wider wall-clock budget and one more |
| 1651 | # worker, because leg 2 is the protocol's only research pass. Shallow-tier |
| 1652 | # bundles keep the one-shot quick constants. Both tiers flow through |
| 1653 | # enrich_nominations' PARAMETERS - the constants themselves are never edited, |
| 1654 | # so neither tier can leak into the other path. |
| 1655 | RESUME_DEEP_ENRICH_DEPTH = "default" |
| 1656 | RESUME_DEEP_ENRICH_MAX_WORKERS = 4 |
| 1657 | RESUME_DEEP_ENRICH_BUDGET_SECONDS = 450.0 |
| 1658 | |
| 1659 | |
| 1660 | def _resume_enrich_budget_seconds(config: dict[str, Any]) -> float: |
| 1661 | """Deep-tier batch budget: LAST30DAYS_ENRICH_BUDGET_SECONDS from the |
| 1662 | RESOLVED config dict only (env.get_config already layers the process env |
| 1663 | over the .env files) - never read from bare os.environ. Blank, |
| 1664 | non-numeric, or non-positive values fall back to the 450s default.""" |
| 1665 | raw = config.get("LAST30DAYS_ENRICH_BUDGET_SECONDS") |
| 1666 | if raw is None or str(raw).strip() == "": |
| 1667 | return RESUME_DEEP_ENRICH_BUDGET_SECONDS |
| 1668 | try: |
| 1669 | value = float(raw) |
| 1670 | except (TypeError, ValueError): |
| 1671 | return RESUME_DEEP_ENRICH_BUDGET_SECONDS |
| 1672 | return value if value > 0 else RESUME_DEEP_ENRICH_BUDGET_SECONDS |
| 1673 | |
| 1674 | |
| 1675 | @dataclass(frozen=True) |
| 1676 | class DiscoverResumeResult: |
| 1677 | """Leg 2 output of the host-judged discovery protocol: the floored, |
| 1678 | folded, velocity-ranked report plus the per-topic angle inputs (keyed by |
| 1679 | surviving nomination id) that the host writes leg-3 angles from. |
| 1680 | ``report.source_status`` is the bundle's restored leg-1 sweep status - |
| 1681 | leg 2 never re-sweeps the listing feeds, so the sweep's degraded-coverage |
| 1682 | signal must survive the handoff instead of reading as clean.""" |
| 1683 | |
| 1684 | report: schema.DiscoveryReport |
| 1685 | angle_inputs: dict[str, dict[str, str]] |
| 1686 | |
| 1687 | |
| 1688 | def run_discover_resume( |
| 1689 | bundle: Any, |
| 1690 | judgments: dict[str, Any], |
| 1691 | *, |
| 1692 | config: dict[str, Any], |
| 1693 | mock: bool = False, |
| 1694 | ) -> DiscoverResumeResult: |
| 1695 | """Protocol leg 2: apply host judgments to the leg-1 bundle, enrich the |
| 1696 | slot winners, and floor/fold/rank on the same code path as the one-shot |
| 1697 | run. |
| 1698 | |
| 1699 | ``bundle`` is a ``discovery_handoff.NominationsBundle`` and ``judgments`` |
| 1700 | the mapping ``discovery_handoff.read_judgments`` returns (annotated |
| 1701 | loosely because discovery_handoff imports this module at load time). |
| 1702 | |
| 1703 | Judgment application is per field: an absent host name falls back to the |
| 1704 | bundle's heuristic name, an absent junk flag to the heuristic junk flag, |
| 1705 | and absent worthiness to the neutral blend default (None -> 50 inside |
| 1706 | ``rerank.judge_blended_score`` - the same treatment the judge-absent path |
| 1707 | always used). Applied names are collision-resolved over the whole pool |
| 1708 | before anything keys on them, and the applied name IS the enrichment |
| 1709 | sub-run topic. |
| 1710 | |
| 1711 | Slot selection: host-junk rows never contend for enrichment slots, and a |
| 1712 | heuristic-junk fallback row with fewer than ``rerank.FLOOR_MIN_SOURCES`` |
| 1713 | distinct seed sources is skipped pre-enrichment (it structurally cannot |
| 1714 | pass the floor's seed-corroboration rule). Both stay eligible to be the |
| 1715 | junk-tracked weak signal of a nothing-solid brief, and the brief prefers |
| 1716 | a non-junk weak signal exactly like the one-shot path. At the floor, |
| 1717 | host-judged rows pass ``junk_shape=False`` (host-junk never earned a |
| 1718 | slot) while heuristic-fallback rows keep their heuristic flag with the |
| 1719 | existing seed-source corroboration. |
| 1720 | |
| 1721 | Velocity, momentum, and the enrichment window all score against the |
| 1722 | bundle's momentum window (from_date/to_date), never the resume-time |
| 1723 | clock: the host may judge up to the handoff TTL after the sweep, and the |
| 1724 | numbers must describe the window the sweep captured. |
| 1725 | """ |
| 1726 | # Runtime-only import: discovery_handoff imports pipeline at module load, |
| 1727 | # so the reverse import must happen at call time (no import-time cycle). |
| 1728 | from . import discovery_handoff |
| 1729 | |
| 1730 | to_date = bundle.to_date |
| 1731 | verdicts = [ |
| 1732 | discovery_handoff.judgment_for(judgments, entry.nomination_id) |
| 1733 | for entry in bundle.nominations |
| 1734 | ] |
| 1735 | applied_names = discovery_handoff.resolve_name_collisions([ |
| 1736 | ( |
| 1737 | entry.nomination, |
| 1738 | verdict.name or entry.heuristic_name or entry.nomination.name, |
| 1739 | ) |
| 1740 | for entry, verdict in zip(bundle.nominations, verdicts) |
| 1741 | ]) |
| 1742 | |
| 1743 | ranked: list[tuple[float, str, Nomination]] = [] |
| 1744 | junk_weak_signal: tuple[float, str] | None = None |
| 1745 | for entry, verdict, name in zip(bundle.nominations, verdicts, applied_names): |
| 1746 | items = entry.nomination.items |
| 1747 | velocity = rerank.discovery_velocity_score(items, as_of_date=to_date) |
| 1748 | seed_source_count = len({item.source for item in items}) |
| 1749 | host_junk = verdict.junk is True |
| 1750 | fallback_junk = verdict.junk is None and entry.heuristic_junk |
| 1751 | if host_junk or ( |
| 1752 | fallback_junk and seed_source_count < rerank.FLOOR_MIN_SOURCES |
| 1753 | ): |
| 1754 | if junk_weak_signal is None or velocity > junk_weak_signal[0]: |
| 1755 | junk_weak_signal = (velocity, name) |
| 1756 | continue |
| 1757 | worthiness = ( |
| 1758 | float(verdict.worthiness) if verdict.worthiness is not None else None |
| 1759 | ) |
| 1760 | blended = rerank.judge_blended_score(velocity, worthiness) |
| 1761 | ranked.append(( |
| 1762 | blended, |
| 1763 | entry.nomination_id, |
| 1764 | replace( |
| 1765 | entry.nomination, |
| 1766 | name=name, |
| 1767 | seed_score=blended, |
| 1768 | junk_shape=( |
| 1769 | False if verdict.junk is not None else entry.heuristic_junk |
| 1770 | ), |
| 1771 | worthiness=worthiness, |
| 1772 | ), |
| 1773 | )) |
| 1774 | |
| 1775 | ranked.sort(key=lambda row: (-row[0], row[2].name.lower())) |
| 1776 | selected = ranked[:ENRICH_LIMIT] |
| 1777 | nominations = [nomination for _blended, _nomination_id, nomination in selected] |
| 1778 | |
| 1779 | if bundle.tier == "shallow": |
| 1780 | depth, max_workers, budget_seconds = ( |
| 1781 | ENRICH_DEPTH, ENRICH_MAX_WORKERS, ENRICH_BUDGET_SECONDS, |
| 1782 | ) |
| 1783 | else: |
| 1784 | depth = RESUME_DEEP_ENRICH_DEPTH |
| 1785 | max_workers = RESUME_DEEP_ENRICH_MAX_WORKERS |
| 1786 | budget_seconds = _resume_enrich_budget_seconds(config) |
| 1787 | |
| 1788 | enriched_entries = enrich_nominations( |
| 1789 | nominations, |
| 1790 | config=config, |
| 1791 | requested_sources=bundle.enrichment_source_boundary, |
| 1792 | mock=mock, |
| 1793 | depth=depth, |
| 1794 | lookback_days=bundle.lookback_days, |
| 1795 | as_of_date=to_date, |
| 1796 | max_workers=max_workers, |
| 1797 | budget_seconds=budget_seconds, |
| 1798 | ) if nominations else [] |
| 1799 | |
| 1800 | # topic_limit mirrors the one-shot default cap (limit=10); the slot cut |
| 1801 | # above already bounds the pool at ENRICH_LIMIT. |
| 1802 | survivors, weak_signal, floor_junk_weak_signal = _floor_survivor_records( |
| 1803 | enriched_entries, to_date=to_date, topic_limit=10, |
| 1804 | ) |
| 1805 | if floor_junk_weak_signal is not None and ( |
| 1806 | junk_weak_signal is None |
| 1807 | or floor_junk_weak_signal[0] > junk_weak_signal[0] |
| 1808 | ): |
| 1809 | junk_weak_signal = floor_junk_weak_signal |
| 1810 | folded = _fold_same_story_records(survivors) |
| 1811 | topics = _records_to_discovery_topics(folded) |
| 1812 | |
| 1813 | nomination_id_by_name = { |
| 1814 | nomination.name: nomination_id |
| 1815 | for _blended, nomination_id, nomination in selected |
| 1816 | } |
| 1817 | angle_inputs = { |
| 1818 | nomination_id_by_name[record["name"]]: { |
| 1819 | "name": record["name"], |
| 1820 | "titles": record["titles"], |
| 1821 | "top_comment": record["top_comment"] or "", |
| 1822 | "engagement": record["engagement_phrase"], |
| 1823 | } |
| 1824 | for record in folded |
| 1825 | } |
| 1826 | |
| 1827 | if weak_signal is None: |
| 1828 | weak_signal = junk_weak_signal |
| 1829 | outcome = "ok" if topics else "nothing-solid" |
| 1830 | plan = schema.DiscoveryPlan( |
| 1831 | domain=bundle.domain, |
| 1832 | category=None, |
| 1833 | subreddits=[], |
| 1834 | sources=( |
| 1835 | list(bundle.requested_sources) |
| 1836 | if bundle.requested_sources |
| 1837 | else sorted({ |
| 1838 | item.source |
| 1839 | for entry in bundle.nominations |
| 1840 | for item in entry.nomination.items |
| 1841 | }) |
| 1842 | ), |
| 1843 | ) |
| 1844 | # The bundle's restored leg-1 sweep status (empty for pre-field bundles): |
| 1845 | # degraded sweep coverage must reach this report's status map and its |
| 1846 | # degraded-sources warning exactly as the one-shot reports it. |
| 1847 | source_status = dict(getattr(bundle, "source_status", None) or {}) |
| 1848 | report = schema.DiscoveryReport( |
| 1849 | domain=bundle.domain, |
| 1850 | range_from=bundle.from_date, |
| 1851 | range_to=to_date, |
| 1852 | generated_at=datetime.now(timezone.utc).isoformat(), |
| 1853 | plan=plan, |
| 1854 | topics=topics, |
| 1855 | source_status=source_status, |
| 1856 | warnings=_discovery_report_warnings(topics, outcome, source_status), |
| 1857 | outcome=outcome, |
| 1858 | weak_signal=weak_signal[1] if weak_signal and not topics else None, |
| 1859 | ) |
| 1860 | return DiscoverResumeResult(report=report, angle_inputs=angle_inputs) |
| 1861 | |
| 1862 | |
| 1863 | def diagnose( |
| 1864 | config: dict[str, Any], |
| 1865 | requested_sources: list[str] | None = None, |
| 1866 | *, |
| 1867 | safe: bool = False, |
| 1868 | x_envelope: bool = False, |
| 1869 | ) -> dict[str, Any]: |
| 1870 | # ``x_envelope`` is True when a validated --x-posts envelope is present for |
| 1871 | # this invocation, so available_sources lists x even without a backend |
| 1872 | # and the optional-source omission note does not fire. |
| 1873 | requested_sources = normalize_requested_sources(requested_sources) |
| 1874 | google_key = _google_key(config) |
| 1875 | x_status = env.get_x_source_status(config, probe=not safe) |
| 1876 | # Compute once and reuse for both the diag flag and available_sources below. |
| 1877 | # safe=True (doctor/--diagnose/--preflight) must stay network-free. |
| 1878 | x_pending = env.x_pending_browser_auth(config, local_only=safe) |
| 1879 | native_web_backend = None |
| 1880 | if config.get("BRAVE_API_KEY"): |
| 1881 | native_web_backend = "brave" |
| 1882 | elif config.get("EXA_API_KEY"): |
| 1883 | native_web_backend = "exa" |
| 1884 | elif config.get("SERPER_API_KEY"): |
| 1885 | native_web_backend = "serper" |
| 1886 | elif config.get("PARALLEL_API_KEY"): |
| 1887 | native_web_backend = "parallel" |
| 1888 | providers_status = { |
| 1889 | "google": bool(google_key), |
| 1890 | "openai": bool(config.get("OPENAI_API_KEY")) and config.get("OPENAI_AUTH_STATUS") == env.AUTH_STATUS_OK, |
| 1891 | "xai": bool(config.get("XAI_API_KEY")), |
| 1892 | "openrouter": bool(config.get("OPENROUTER_API_KEY")), |
| 1893 | "perplexity": bool(config.get("PERPLEXITY_API_KEY")), |
| 1894 | } |
| 1895 | reasoning_provider_available = any( |
| 1896 | providers_status[name] for name in ("google", "openai", "xai", "openrouter") |
| 1897 | ) |
| 1898 | external_commands = { |
| 1899 | "yt-dlp": bool(which("yt-dlp")), |
| 1900 | "digg-pp-cli": bool(which("digg-pp-cli")), |
| 1901 | "arxiv-pp-cli": bool(which("arxiv-pp-cli")), |
| 1902 | "techmeme-pp-cli": bool(which("techmeme-pp-cli")), |
| 1903 | "trustpilot-pp-cli": bool(which("trustpilot-pp-cli")), |
| 1904 | "brightdata": bool(which("brightdata")), |
| 1905 | "gh": bool(which("gh")), |
| 1906 | } |
| 1907 | # Network-free two-field probe (bird_installed/bird_authenticated |
| 1908 | # precedent): "installed" is PATH resolution, "authenticated" is a |
| 1909 | # presence-only credential signal that never reads the secret. |
| 1910 | brightdata_status = brightdata.gate_status(config) |
| 1911 | credential_destinations = { |
| 1912 | "global_env": str(env.CONFIG_FILE) if env.CONFIG_FILE else None, |
| 1913 | } |
| 1914 | browser_cookies = { |
| 1915 | "mode": config.get("_BROWSER_COOKIE_MODE", "off"), |
| 1916 | "browsers": list(config.get("_BROWSER_COOKIE_BROWSERS") or []), |
| 1917 | "reads_values": False if safe else config.get("_BROWSER_COOKIE_MODE") == "read", |
| 1918 | } |
| 1919 | ignored_project_keys = list(config.get("_IGNORED_PROJECT_CONFIG_KEYS") or []) |
| 1920 | ignored_endpoint_overrides = [ |
| 1921 | key for key in ignored_project_keys if key in permission_preflight.ENDPOINT_OVERRIDE_KEYS |
| 1922 | ] |
| 1923 | local_writes: list[dict[str, str]] = [] |
| 1924 | if config.get("LAST30DAYS_MEMORY_DIR"): |
| 1925 | local_writes.append({"kind": "report", "path": str(config.get("LAST30DAYS_MEMORY_DIR"))}) |
| 1926 | diag = { |
| 1927 | "providers": providers_status, |
| 1928 | "local_mode": not reasoning_provider_available, |
| 1929 | "reasoning_provider": (config.get("LAST30DAYS_REASONING_PROVIDER") or "auto").lower(), |
| 1930 | # The host-fetched connector lane serves X when no engine backend |
| 1931 | # exists and the model declared the lane. |
| 1932 | "x_backend": x_status["source"] or ( |
| 1933 | "connector" if env.x_host_lane_declared(config) else None |
| 1934 | ), |
| 1935 | "bird_installed": x_status["bird_installed"], |
| 1936 | "bird_authenticated": x_status["bird_authenticated"], |
| 1937 | "bird_username": x_status["bird_username"], |
| 1938 | "x_pending_browser_auth": x_pending, |
| 1939 | "xquik_available": x_status.get("xquik_available", False), |
| 1940 | "xquik_working": x_status.get("xquik_working"), |
| 1941 | "xquik_status": x_status.get("xquik_status", ""), |
| 1942 | "native_web_backend": native_web_backend, |
| 1943 | "native_search": env.is_native_search(config), |
| 1944 | "has_scrapecreators": bool(config.get("SCRAPECREATORS_API_KEY")), |
| 1945 | "has_github": bool(config.get("GITHUB_TOKEN") or which("gh")), |
| 1946 | "brightdata_installed": brightdata_status["brightdata_installed"], |
| 1947 | "brightdata_authenticated": brightdata_status["brightdata_authenticated"], |
| 1948 | # safe=True (doctor/--diagnose/--preflight) must stay network-free: |
| 1949 | # answer X availability from local evidence only. x_pending is |
| 1950 | # precomputed by diagnose() to avoid double evaluation. |
| 1951 | "available_sources": available_sources( |
| 1952 | config, requested_sources, x_pending=x_pending, local_only=safe, |
| 1953 | x_envelope=x_envelope, |
| 1954 | ), |
| 1955 | "safe": safe, |
| 1956 | "config_source": config.get("_CONFIG_SOURCE"), |
| 1957 | "ignored_project_config": config.get("_IGNORED_PROJECT_CONFIG"), |
| 1958 | "ignored_project_config_keys": ignored_project_keys, |
| 1959 | "ignored_endpoint_overrides": ignored_endpoint_overrides, |
| 1960 | "browser_cookies": browser_cookies, |
| 1961 | "external_commands": external_commands, |
| 1962 | "credential_destinations": credential_destinations, |
| 1963 | "local_writes": local_writes, |
| 1964 | } |
| 1965 | diag["permission_preflight"] = permission_preflight.build(config, diag) |
| 1966 | return diag |
| 1967 | |
| 1968 | |
| 1969 | def _inner_max_workers(stream_count: int, *, internal_subrun: bool) -> int: |
| 1970 | """Worker-pool size for the per-stream fanout inside a single pipeline run. |
| 1971 | |
| 1972 | Top-level runs use up to 16 workers. Subruns of ``run_competitor_fanout`` |
| 1973 | cap the inner pool to 4 so a six-way competitor fan-out stays below |
| 1974 | roughly 30 worker threads in aggregate instead of ~96. |
| 1975 | """ |
| 1976 | if internal_subrun: |
| 1977 | return max(2, min(4, stream_count or 1)) |
| 1978 | return max(4, min(16, stream_count or 1)) |
| 1979 | |
| 1980 | |
| 1981 | def _load_library_context( |
| 1982 | *, |
| 1983 | topic: str, |
| 1984 | config: dict[str, Any], |
| 1985 | mock: bool, |
| 1986 | internal_subrun: bool, |
| 1987 | x_handle: str | None, |
| 1988 | github_user: str | None, |
| 1989 | github_repos: list[str] | None, |
| 1990 | save_dir: Path | str | None = None, |
| 1991 | ) -> tuple[list[schema.LibraryContext], str | None]: |
| 1992 | """Resolve compact prior-run context without making a research run depend on it.""" |
| 1993 | setting = str(config.get("LAST30DAYS_LIBRARY_CONTEXT") or "off").strip().lower() |
| 1994 | if mock or internal_subrun or setting in {"0", "false", "no", "off"}: |
| 1995 | return [], None |
| 1996 | if save_dir == "": |
| 1997 | return [], None |
| 1998 | |
| 1999 | memory_dir = ( |
| 2000 | save_dir |
| 2001 | if save_dir is not None |
| 2002 | else config.get("LAST30DAYS_MEMORY_DIR") or library.DEFAULT_MEMORY_DIR |
| 2003 | ) |
| 2004 | briefs_dir = config.get("_LAST30DAYS_LIBRARY_BRIEFS_DIR") or ( |
| 2005 | Path(memory_dir).expanduser() / "briefings" |
| 2006 | if save_dir is not None |
| 2007 | else library.DEFAULT_BRIEFS_DIR |
| 2008 | ) |
| 2009 | db_path = config.get("_LAST30DAYS_LIBRARY_DB") |
| 2010 | if not db_path: |
| 2011 | db_path = ( |
| 2012 | Path(memory_dir).expanduser().resolve() / ".last30days-library.db" |
| 2013 | if save_dir is not None |
| 2014 | else library_index.DEFAULT_LIBRARY_DB |
| 2015 | ) |
| 2016 | store_db = config.get("_LAST30DAYS_STORE_DB") |
| 2017 | if not store_db: |
| 2018 | # Scoped runs read only a store inside the save dir (usually absent); |
| 2019 | # the shared store would leak other scopes' sightings into this one. |
| 2020 | store_db = ( |
| 2021 | Path(memory_dir).expanduser().resolve() / "research.db" |
| 2022 | if save_dir is not None |
| 2023 | else library_index.DEFAULT_STORE_DB |
| 2024 | ) |
| 2025 | queries = [topic, x_handle or "", github_user or "", *(github_repos or [])] |
| 2026 | queries = list(dict.fromkeys(value.strip() for value in queries if value and value.strip())) |
| 2027 | try: |
| 2028 | library_index.sync_library(memory_dir, briefs_dir, db_path=db_path) |
| 2029 | matches: list[library_index.LibrarySearchMatch] = [] |
| 2030 | for query_text in queries: |
| 2031 | matches.extend( |
| 2032 | library_index.search( |
| 2033 | query_text, |
| 2034 | limit=6, |
| 2035 | db_path=db_path, |
| 2036 | store_db_path=store_db, |
| 2037 | ) |
| 2038 | ) |
| 2039 | except (library_index.LibrarySearchUnavailable, OSError, sqlite3.DatabaseError) as exc: |
| 2040 | return [], f"Library context unavailable: {exc}" |
| 2041 | |
| 2042 | contexts: list[schema.LibraryContext] = [] |
| 2043 | seen_runs: set[tuple[str, date]] = set() |
| 2044 | for match in sorted( |
| 2045 | matches, |
| 2046 | key=lambda item: (-item.published_date.toordinal(), item.rank, item.topic.casefold()), |
| 2047 | ): |
| 2048 | if match.run_key in seen_runs: |
| 2049 | continue |
| 2050 | seen_runs.add(match.run_key) |
| 2051 | contexts.append( |
| 2052 | schema.LibraryContext( |
| 2053 | topic=match.topic, |
| 2054 | published_date=match.published_date.isoformat(), |
| 2055 | headline=match.headline, |
| 2056 | summary=match.snippet or match.headline, |
| 2057 | source_kind=match.source_kind, |
| 2058 | ) |
| 2059 | ) |
| 2060 | if len(contexts) == 3: |
| 2061 | break |
| 2062 | return contexts, None |
| 2063 | |
| 2064 | |
| 2065 | def run( |
| 2066 | *, |
| 2067 | topic: str, |
| 2068 | config: dict[str, Any], |
| 2069 | depth: str, |
| 2070 | requested_sources: list[str] | None = None, |
| 2071 | mock: bool = False, |
| 2072 | x_handle: str | None = None, |
| 2073 | x_related: list[str] | None = None, |
| 2074 | web_backend: str = "auto", |
| 2075 | external_plan: dict | None = None, |
| 2076 | subreddits: list[str] | None = None, |
| 2077 | tiktok_hashtags: list[str] | None = None, |
| 2078 | tiktok_creators: list[str] | None = None, |
| 2079 | ig_creators: list[str] | None = None, |
| 2080 | lookback_days: int = 30, |
| 2081 | as_of_date: str | None = None, |
| 2082 | github_user: str | None = None, |
| 2083 | github_repos: list[str] | None = None, |
| 2084 | trustpilot_domain: str | None = None, |
| 2085 | trustpilot_domain_is_hint: bool = False, |
| 2086 | hiring_signals_mode: bool = False, |
| 2087 | internal_subrun: bool = False, |
| 2088 | suppress_x_host_lane: bool = False, |
| 2089 | save_dir: Path | str | None = None, |
| 2090 | corpus_dirs: list[str] | None = None, |
| 2091 | corpus_all_time: bool = False, |
| 2092 | x_posts: x_envelope.Envelope | None = None, |
| 2093 | ) -> schema.Report: |
| 2094 | # ``suppress_x_host_lane`` is distinct from ``internal_subrun``: comparison |
| 2095 | # entities share the latter and must still honor the connector lane; |
| 2096 | # only discovery enrichment passes set the former. |
| 2097 | # ``x_posts`` is a validated ``--x-posts`` envelope: when present |
| 2098 | # it replaces the engine's X fetch for this run and is served once. |
| 2099 | # Standalone runs (not competitor/discover sub-runs) own the YouTube |
| 2100 | # search-cache lifecycle. Comparison fan-out clears once before submit so |
| 2101 | # parallel entity sub-runs can still share in-run hits. |
| 2102 | if not internal_subrun: |
| 2103 | youtube_yt.reset_search_cache() |
| 2104 | settings = _resolve_depth_settings(depth, config) |
| 2105 | requested_sources = normalize_requested_sources(requested_sources) |
| 2106 | # Wall-clock origin for budget-aware enrichment lanes. Amazon review |
| 2107 | # enrichment starts at search time (inside _retrieve_stream_impl) so it |
| 2108 | # overlaps other sources instead of waiting for them all to finish. |
| 2109 | run_started = time.monotonic() |
| 2110 | from_date, to_date = dates.get_date_range(lookback_days, as_of_date=as_of_date) |
| 2111 | resolved_corpus_dirs = corpus.resolve_directories( |
| 2112 | corpus_dirs or config.get("_CORPUS_DIRS"), |
| 2113 | config.get("LAST30DAYS_CORPUS_DIRS"), |
| 2114 | ) |
| 2115 | excluded_sources = { |
| 2116 | source.strip().lower() |
| 2117 | for source in str(config.get("EXCLUDE_SOURCES") or "").split(",") |
| 2118 | if source.strip() |
| 2119 | } |
| 2120 | corpus_enabled = bool(resolved_corpus_dirs) and "corpus" not in excluded_sources |
| 2121 | corpus_requested = bool(requested_sources and "corpus" in requested_sources) |
| 2122 | if corpus_enabled and requested_sources and "corpus" not in requested_sources: |
| 2123 | requested_sources = [*requested_sources, "corpus"] |
| 2124 | |
| 2125 | # Host-fetched X lane. EXCLUDE_SOURCES=x or a --search list without |
| 2126 | # x wins: the envelope is ignored with a receipt line and stays unconsumed. |
| 2127 | envelope = x_posts |
| 2128 | if envelope is not None and ( |
| 2129 | "x" in excluded_sources |
| 2130 | or (requested_sources and "x" not in requested_sources) |
| 2131 | ): |
| 2132 | log.source_log( |
| 2133 | "x", "host-fetched X: envelope ignored (x is excluded from this run)", |
| 2134 | tty_only=False, |
| 2135 | ) |
| 2136 | envelope = None |
| 2137 | # The lane signal without an envelope is a broken handoff, not a reason to |
| 2138 | # spend a backup backend: X records the fixed not-passed outcome. |
| 2139 | x_lane_missing = ( |
| 2140 | envelope is None |
| 2141 | and not mock |
| 2142 | and not suppress_x_host_lane |
| 2143 | and env.x_host_lane_declared(config) |
| 2144 | ) |
| 2145 | if envelope is not None or x_lane_missing: |
| 2146 | # Ride the config dict (the _polymarket_keywords idiom) so the stream |
| 2147 | # workers and the handle-lane section see it without widening their |
| 2148 | # signatures. Copy first: comparison entities shallow-copy the shared |
| 2149 | # config and must never inherit another entity's envelope. |
| 2150 | config = dict(config) |
| 2151 | config["_x_envelope"] = envelope |
| 2152 | config["_x_lane_missing"] = x_lane_missing |
| 2153 | |
| 2154 | # Gate StockTwits to ticker/crypto topics. Single chokepoint: when False, |
| 2155 | # available_sources() never registers stocktwits, so the planner can't |
| 2156 | # assign it (eligible_sources = available ∩ capabilities). |
| 2157 | config["_financial_topic"] = stocktwits.is_financial_topic(topic) |
| 2158 | |
| 2159 | if mock: |
| 2160 | runtime = providers.mock_runtime(config, depth) |
| 2161 | reasoning_provider = None |
| 2162 | available = list(requested_sources or MOCK_AVAILABLE_SOURCES) |
| 2163 | if corpus_enabled and "corpus" not in available: |
| 2164 | available.append("corpus") |
| 2165 | if not corpus_enabled and not corpus_requested: |
| 2166 | available = [source for source in available if source != "corpus"] |
| 2167 | if not requested_sources and not hiring_signals_mode and not _company_topic_likely(topic): |
| 2168 | available = [source for source in available if source != "jobs"] |
| 2169 | else: |
| 2170 | runtime, reasoning_provider = providers.resolve_runtime(config, depth) |
| 2171 | available = available_sources( |
| 2172 | config, requested_sources, |
| 2173 | suppress_x_host_lane=suppress_x_host_lane, |
| 2174 | x_envelope=envelope is not None, |
| 2175 | ) |
| 2176 | if requested_sources: |
| 2177 | available = [source for source in available if source in requested_sources] |
| 2178 | # Keep an explicitly requested but unconfigured corpus in the plan long |
| 2179 | # enough to record its skipped-unconfigured source outcome. It is never |
| 2180 | # submitted to the network executor below. |
| 2181 | if corpus_requested and "corpus" not in excluded_sources and "corpus" not in available: |
| 2182 | available.append("corpus") |
| 2183 | if web_backend == "none": |
| 2184 | available = [s for s in available if s != "grounding"] |
| 2185 | elif web_backend in ("brave", "exa", "serper", "parallel", "parallel-mcp", "keyless") and "grounding" not in available: |
| 2186 | available.append("grounding") |
| 2187 | if ( |
| 2188 | hiring_signals_mode |
| 2189 | or (not requested_sources and _company_topic_likely(topic)) |
| 2190 | ) and "jobs" not in available: |
| 2191 | available.append("jobs") |
| 2192 | if hiring_signals_mode: |
| 2193 | config = dict(config) |
| 2194 | config["_hiring_signals_mode"] = True |
| 2195 | if not requested_sources: |
| 2196 | available = ["jobs"] |
| 2197 | if not available: |
| 2198 | raise RuntimeError("No sources are available for this run.") |
| 2199 | |
| 2200 | planner_requested_sources = requested_sources |
| 2201 | if hiring_signals_mode and not planner_requested_sources: |
| 2202 | planner_requested_sources = ["jobs"] |
| 2203 | |
| 2204 | if external_plan is not None: |
| 2205 | # External plan provided (e.g., from Claude Code via --plan flag). |
| 2206 | # Explicit input is a contract: validate it before permissive sanitization. |
| 2207 | planner.validate_external_plan(external_plan) |
| 2208 | plan = planner._sanitize_plan( |
| 2209 | external_plan, topic, available, planner_requested_sources, depth, |
| 2210 | honor_plan_sources=True, |
| 2211 | ) |
| 2212 | plan_source = "external" |
| 2213 | else: |
| 2214 | plan = planner.plan_query( |
| 2215 | topic=topic, |
| 2216 | available_sources=available, |
| 2217 | requested_sources=planner_requested_sources, |
| 2218 | depth=depth, |
| 2219 | provider=None if mock else reasoning_provider, |
| 2220 | model=None if mock else runtime.planner_model, |
| 2221 | context=config.get("_auto_resolve_context", ""), |
| 2222 | internal_subrun=internal_subrun, |
| 2223 | ) |
| 2224 | # Source labelling: the fallback path annotates notes with "fallback-plan" |
| 2225 | # or "deterministic-comparison-plan"; anything else came from the LLM. |
| 2226 | if any("fallback" in note or "deterministic" in note for note in (plan.notes or [])): |
| 2227 | plan_source = "deterministic" |
| 2228 | elif not mock and reasoning_provider and runtime.planner_model: |
| 2229 | plan_source = "llm" |
| 2230 | else: |
| 2231 | plan_source = "deterministic" |
| 2232 | |
| 2233 | # Safety net: ensure grounding appears in all subqueries even if the planner |
| 2234 | # omits it. This is redundant when the planner includes grounding via |
| 2235 | # SOURCE_CAPABILITIES, but kept as a fallback. |
| 2236 | if ( |
| 2237 | web_backend != "none" |
| 2238 | and "grounding" in available |
| 2239 | and "drill-mode" not in plan.notes |
| 2240 | ): |
| 2241 | for sq in plan.subqueries: |
| 2242 | if "grounding" not in sq.sources: |
| 2243 | sq.sources.append("grounding") |
| 2244 | if "drill-mode" not in plan.notes: |
| 2245 | # Drill plans re-fetch only the sources that contributed to the matched |
| 2246 | # cluster; the company-topic jobs injection must not widen that set. |
| 2247 | _ensure_jobs_in_plan(plan, available, explicit=hiring_signals_mode, topic=topic) |
| 2248 | if "corpus" in available and plan.subqueries: |
| 2249 | # Corpus is deterministic and user-registered, so it always gets one |
| 2250 | # bounded stream even when a quick/LLM plan omits it. Reuse the primary |
| 2251 | # subquery instead of multiplying local scans across every subquery. |
| 2252 | if "corpus" not in plan.subqueries[0].sources: |
| 2253 | plan.subqueries[0].sources.append("corpus") |
| 2254 | if "corpus" not in plan.source_weights: |
| 2255 | plan.source_weights["corpus"] = 1.0 |
| 2256 | plan.source_weights = planner._normalize_weights(plan.source_weights) |
| 2257 | |
| 2258 | # Add the paid-only Perplexity lane after all normal-source safety nets. |
| 2259 | # This preserves the planner's primary subquery, gives the bounded paid |
| 2260 | # call the whole user topic, and prevents grounding, jobs, or corpus from |
| 2261 | # being attached to the dedicated lane. |
| 2262 | _ensure_perplexity_in_plan( |
| 2263 | plan, |
| 2264 | topic, |
| 2265 | available, |
| 2266 | force=bool(config.get("_deep_research")), |
| 2267 | ) |
| 2268 | |
| 2269 | # Always-on planner trace. Emits one summary line plus one per subquery |
| 2270 | # so retrieval-breadth failures like the 2026-04-19 Hermes Agent Use Cases |
| 2271 | # disaster are visible without --debug. Stderr only; does not leak into |
| 2272 | # the user-facing stdout synthesis. |
| 2273 | print( |
| 2274 | f"[Planner] Plan: intent={plan.intent}, freshness={plan.freshness_mode}, " |
| 2275 | f"cluster_mode={plan.cluster_mode}, subqueries={len(plan.subqueries)}, " |
| 2276 | f"source={plan_source}", |
| 2277 | file=sys.stderr, |
| 2278 | ) |
| 2279 | if plan.subqueries: |
| 2280 | for index, sq in enumerate(plan.subqueries, start=1): |
| 2281 | sources_str = ",".join(sq.sources) if sq.sources else "(none)" |
| 2282 | print( |
| 2283 | f"[Planner] sq{index} label={sq.label} " |
| 2284 | f'search="{sq.search_query}" sources=[{sources_str}]', |
| 2285 | file=sys.stderr, |
| 2286 | ) |
| 2287 | else: |
| 2288 | print("[Planner] (no subqueries in plan)", file=sys.stderr) |
| 2289 | |
| 2290 | bundle = schema.RetrievalBundle(artifacts={"grounding": []}) |
| 2291 | if envelope is not None: |
| 2292 | # The footer's X provenance reads "via X connector" (render._render_stats). |
| 2293 | bundle.artifacts["x_provenance"] = "connector" |
| 2294 | # Handles the user named explicitly. Available before any retrieval, unlike |
| 2295 | # the entity-extracted set, so Phase 1 and quick-depth runs get first-party |
| 2296 | # protection too. Without this the exemption reached only the Phase 2 |
| 2297 | # supplement path -- which quick runs skip entirely -- so a subject-authored |
| 2298 | # post retrieved in Phase 1 was still pruned before fusion, which is exactly |
| 2299 | # the evidence loss this change exists to prevent. |
| 2300 | explicit_first_party = { |
| 2301 | h.lstrip("@").strip().lower() |
| 2302 | for h in ([x_handle, github_user, *(x_related or [])]) |
| 2303 | if h and h.strip() |
| 2304 | } |
| 2305 | # Creator accounts named via --ig-creators / --creators carry the same |
| 2306 | # explicit intent: the run is searching those accounts, and a creator's |
| 2307 | # caption rarely repeats the topic's literal tokens, so without an |
| 2308 | # exemption the relevance floor prunes them as third-party noise (issue |
| 2309 | # #1101: 36 creator reels fetched, 0 reported). The exemption is scoped |
| 2310 | # per platform, NOT merged into the global set: each flag names accounts |
| 2311 | # on one platform, and an unrelated same-name account elsewhere must not |
| 2312 | # bypass the floors. |
| 2313 | creator_first_party = _creator_first_party_by_source(tiktok_creators, ig_creators) |
| 2314 | # Real X handles: --x-handle, --x-related, or @mentions in the topic. These |
| 2315 | # determine whether the deferred X floor applies. Topic words like "peter" |
| 2316 | # are NOT real handles and should not trigger the floor — when no real |
| 2317 | # handle is identified, the floor is skipped entirely (policy: a noisier |
| 2318 | # report beats losing the subject's evidence). |
| 2319 | explicit_x_handles = { |
| 2320 | h.lstrip("@").strip().lower() |
| 2321 | for h in ([x_handle, *(x_related or [])]) |
| 2322 | if h and h.strip() |
| 2323 | } | _topic_handle_mentions(topic) |
| 2324 | # Plus handle-shaped tokens from the topic. Phase 1 and quick-depth runs |
| 2325 | # never reach automatic handle resolution, so without this a quick search |
| 2326 | # naming a subject still discards everything that subject wrote. |
| 2327 | explicit_first_party |= _topic_first_party_candidates(topic) |
| 2328 | |
| 2329 | for source in (requested_sources or []): |
| 2330 | if source not in available: |
| 2331 | bundle.record_failure( |
| 2332 | source, |
| 2333 | schema.SKIPPED_UNCONFIGURED, |
| 2334 | "Source was requested but is not configured for this run.", |
| 2335 | attempted=False, |
| 2336 | ) |
| 2337 | if corpus_requested and not corpus_enabled: |
| 2338 | bundle.record_failure( |
| 2339 | "corpus", |
| 2340 | schema.SKIPPED_UNCONFIGURED, |
| 2341 | "Corpus was requested but no readable directory was configured.", |
| 2342 | attempted=False, |
| 2343 | ) |
| 2344 | # Expose plan_source to the renderer so render_compact can emit the |
| 2345 | # DEGRADED RUN banner when a named-entity topic was invoked bare |
| 2346 | # (source=deterministic AND no pre-research flags). LAW 7 backstop. |
| 2347 | bundle.artifacts["plan_source"] = plan_source |
| 2348 | bundle.artifacts["corpus_in_export"] = bool(config.get("_CORPUS_IN_EXPORT")) |
| 2349 | # Hiring-signals is deliberately jobs-only with no multi-source --plan, so |
| 2350 | # the LAW 7 degraded-run and Step 0.55 pre-research banners do not apply - |
| 2351 | # they would contradict the documented jobs-scoped flow. Suppress them. |
| 2352 | bundle.artifacts["hiring_signals_mode"] = hiring_signals_mode |
| 2353 | # Record the resolved Amazon keyword whenever the lane is active, so the |
| 2354 | # footer can name it on an empty result. A search that matched nothing |
| 2355 | # still spent a credit, and the fix is almost always the keyword -- a |
| 2356 | # suppressed line means nobody ever learns it was wrong. |
| 2357 | if "amazon" in (available or []): |
| 2358 | bundle.artifacts["amazon_query"] = ( |
| 2359 | str(config.get("_amazon_query") or "").strip() or topic |
| 2360 | ) |
| 2361 | |
| 2362 | # Project-mode or person-mode GitHub: run once before the main subquery loop |
| 2363 | _github_custom_done = False |
| 2364 | _github_enriched_repos: set[str] = set() |
| 2365 | |
| 2366 | # Project mode takes priority over person mode |
| 2367 | if github_repos and "github" in available: |
| 2368 | bundle.mark_attempted("github") |
| 2369 | try: |
| 2370 | project_items = github.search_github_project( |
| 2371 | github_repos, from_date, to_date, |
| 2372 | depth=depth, token=config.get("GITHUB_TOKEN"), |
| 2373 | ) |
| 2374 | if project_items: |
| 2375 | normalized = _normalize_score_dedupe( |
| 2376 | "github", project_items, from_date, to_date, |
| 2377 | freshness_mode=plan.freshness_mode, |
| 2378 | ranking_query=f"What are {', '.join(github_repos)} doing on GitHub?", |
| 2379 | ) |
| 2380 | primary_label = plan.subqueries[0].label if plan.subqueries else "primary" |
| 2381 | bundle.add_items(primary_label, "github", normalized) |
| 2382 | _github_custom_done = True |
| 2383 | _github_enriched_repos = {r.lower() for r in github_repos} |
| 2384 | except Exception as exc: |
| 2385 | bundle.errors_by_source["github"] = f"Project-mode failed: {exc}" |
| 2386 | state, attempted = _classify_source_failure(exc) |
| 2387 | bundle.record_failure("github", state, str(exc), attempted=attempted) |
| 2388 | |
| 2389 | _github_person_done = False |
| 2390 | if github_user and "github" in available and not _github_custom_done: |
| 2391 | bundle.mark_attempted("github") |
| 2392 | _github_person_done = True |
| 2393 | try: |
| 2394 | person_items = github.search_github_person( |
| 2395 | github_user, from_date, to_date, |
| 2396 | depth=depth, token=config.get("GITHUB_TOKEN"), |
| 2397 | ) |
| 2398 | if person_items: |
| 2399 | normalized = _normalize_score_dedupe( |
| 2400 | "github", person_items, from_date, to_date, |
| 2401 | freshness_mode=plan.freshness_mode, |
| 2402 | ranking_query=f"What is @{github_user} doing on GitHub?", |
| 2403 | ) |
| 2404 | # Use the first subquery's label so RRF can look up the weight |
| 2405 | primary_label = plan.subqueries[0].label if plan.subqueries else "primary" |
| 2406 | bundle.add_items(primary_label, "github", normalized) |
| 2407 | else: |
| 2408 | # A pinned --github-user that yields nothing must not be |
| 2409 | # silently backfilled by generic keyword search: the report |
| 2410 | # would then present unrelated repos as this person's work. |
| 2411 | bundle.record_failure( |
| 2412 | "github", |
| 2413 | "no-results", |
| 2414 | f"Person mode found no activity for @{github_user} in the window", |
| 2415 | ) |
| 2416 | except Exception as exc: |
| 2417 | bundle.errors_by_source["github"] = f"Person-mode failed: {exc}" |
| 2418 | state, attempted = _classify_source_failure(exc) |
| 2419 | bundle.record_failure("github", state, str(exc), attempted=attempted) |
| 2420 | |
| 2421 | # Trustpilot session warm-up happens inside search_trustpilot at the |
| 2422 | # first (capped, single) fetch -- lazily, so it never delays the other |
| 2423 | # sources' streams and never fires for runs whose plan fetches no |
| 2424 | # Trustpilot. The module-level lock in lib/trustpilot.py serializes |
| 2425 | # concurrent vs-mode sub-runs so they never race Chrome harvests. |
| 2426 | |
| 2427 | # Thread-safe set prevents redundant fetches after a source returns 429 |
| 2428 | rate_limited_sources: set[str] = set() |
| 2429 | rate_limit_lock = threading.Lock() |
| 2430 | |
| 2431 | # Local corpus retrieval is intentionally outside the network executor and |
| 2432 | # retry budget. One bounded stream participates in the same signal scoring, |
| 2433 | # fusion, reranking, and per-source result cap as remote sources. |
| 2434 | if corpus_enabled and plan.subqueries: |
| 2435 | primary = plan.subqueries[0] |
| 2436 | bundle.mark_attempted("corpus") |
| 2437 | result = corpus.search( |
| 2438 | topic, |
| 2439 | resolved_corpus_dirs, |
| 2440 | from_date=from_date, |
| 2441 | to_date=to_date, |
| 2442 | all_time=corpus_all_time, |
| 2443 | limit=settings["per_stream_limit"], |
| 2444 | cache_dir=env.CONFIG_DIR, |
| 2445 | ) |
| 2446 | prepared_query = relevance.PreparedQuery(primary.ranking_query) |
| 2447 | lookback_window_days = ( |
| 2448 | datetime.strptime(to_date, "%Y-%m-%d").date() |
| 2449 | - datetime.strptime(from_date, "%Y-%m-%d").date() |
| 2450 | ).days |
| 2451 | corpus_items = signals.annotate_stream( |
| 2452 | result.items, |
| 2453 | prepared_query, |
| 2454 | plan.freshness_mode, |
| 2455 | reference_date=to_date, |
| 2456 | max_days=lookback_window_days, |
| 2457 | ) |
| 2458 | corpus_items = signals.prune_low_relevance(corpus_items) |
| 2459 | corpus_items = dedupe.dedupe_items(corpus_items) |
| 2460 | for item in corpus_items: |
| 2461 | item.snippet = snippet.extract_best_snippet(item, prepared_query) |
| 2462 | bundle.add_items(primary.label, "corpus", corpus_items) |
| 2463 | if result.notes: |
| 2464 | outcome = bundle.source_status["corpus"] |
| 2465 | bundle.source_status["corpus"] = schema.SourceOutcome( |
| 2466 | source="corpus", |
| 2467 | state=outcome.state, |
| 2468 | items_returned=outcome.items_returned, |
| 2469 | attempted=True, |
| 2470 | detail="; ".join(result.notes), |
| 2471 | ) |
| 2472 | bundle.artifacts["corpus"] = { |
| 2473 | "files_scanned": result.files_scanned, |
| 2474 | "cache_hits": result.cache_hits, |
| 2475 | "all_time": corpus_all_time, |
| 2476 | } |
| 2477 | |
| 2478 | futures = {} |
| 2479 | # Per-source fetch budget prevents redundant API calls |
| 2480 | source_fetch_count: dict[str, int] = {} |
| 2481 | stream_count = sum( |
| 2482 | 1 |
| 2483 | for subquery in plan.subqueries |
| 2484 | for source in subquery.sources |
| 2485 | if source in available and source != "corpus" |
| 2486 | ) |
| 2487 | max_workers = _inner_max_workers(stream_count, internal_subrun=internal_subrun) |
| 2488 | with ThreadPoolExecutor(max_workers=max_workers) as executor: |
| 2489 | for subquery in plan.subqueries: |
| 2490 | for source in subquery.sources: |
| 2491 | if source not in available: |
| 2492 | continue |
| 2493 | if source == "corpus": |
| 2494 | continue |
| 2495 | # Skip GitHub keyword search if person-mode already ran |
| 2496 | if source == "github" and (_github_person_done or _github_custom_done): |
| 2497 | continue |
| 2498 | # Enforce per-source fetch cap. A CLI override (issue #716) raises |
| 2499 | # the cap for capped sources so every X subquery in a multi-angle |
| 2500 | # --plan fetches, instead of only the first two. |
| 2501 | cap = _source_fetch_cap(source, config) |
| 2502 | if cap is not None: |
| 2503 | if cap <= 0: |
| 2504 | continue |
| 2505 | current = source_fetch_count.get(source, 0) |
| 2506 | if current >= cap: |
| 2507 | continue |
| 2508 | shared_paid_budget = config.get("_perplexity_paid_budget") |
| 2509 | if ( |
| 2510 | source == "perplexity" |
| 2511 | and isinstance(shared_paid_budget, PaidSourceBudget) |
| 2512 | and not shared_paid_budget.try_consume( |
| 2513 | cap, |
| 2514 | claimant=topic, |
| 2515 | ) |
| 2516 | ): |
| 2517 | bundle.artifacts.setdefault("paid_source_budget", {})[ |
| 2518 | "perplexity" |
| 2519 | ] = { |
| 2520 | "state": "skipped-budget", |
| 2521 | "attempted": False, |
| 2522 | "owner": shared_paid_budget.owner, |
| 2523 | "claimant": topic, |
| 2524 | } |
| 2525 | continue |
| 2526 | source_fetch_count[source] = current + 1 |
| 2527 | bundle.mark_attempted(source) |
| 2528 | futures[ |
| 2529 | executor.submit( |
| 2530 | _retrieve_stream, |
| 2531 | topic=topic, |
| 2532 | subquery=subquery, |
| 2533 | source=source, |
| 2534 | config=config, |
| 2535 | depth=depth, |
| 2536 | date_range=(from_date, to_date), |
| 2537 | runtime=runtime, |
| 2538 | mock=mock, |
| 2539 | rate_limited_sources=rate_limited_sources, |
| 2540 | rate_limit_lock=rate_limit_lock, |
| 2541 | web_backend=web_backend, |
| 2542 | raw_topic=topic, |
| 2543 | subreddits=subreddits, |
| 2544 | tiktok_hashtags=tiktok_hashtags, |
| 2545 | tiktok_creators=tiktok_creators, |
| 2546 | ig_creators=ig_creators, |
| 2547 | trustpilot_domain=trustpilot_domain, |
| 2548 | trustpilot_domain_is_hint=trustpilot_domain_is_hint, |
| 2549 | run_started=run_started, |
| 2550 | ) |
| 2551 | ] = (subquery, source) |
| 2552 | |
| 2553 | for future in as_completed(futures): |
| 2554 | subquery, source = futures[future] |
| 2555 | try: |
| 2556 | raw_items, artifact = future.result() |
| 2557 | except Exception as exc: |
| 2558 | # Share 429 signal so pending futures skip this source |
| 2559 | if _is_rate_limit_error(exc): |
| 2560 | with rate_limit_lock: |
| 2561 | rate_limited_sources.add(source) |
| 2562 | bundle.errors_by_source[source] = str(exc) |
| 2563 | state, attempted = _classify_source_failure(exc) |
| 2564 | bundle.record_failure(source, state, str(exc), attempted=attempted) |
| 2565 | continue |
| 2566 | # Retry once for transient 5xx errors |
| 2567 | if _is_transient_error(exc): |
| 2568 | time.sleep(3) |
| 2569 | try: |
| 2570 | raw_items, artifact = _retrieve_stream( |
| 2571 | topic=topic, subquery=subquery, source=source, |
| 2572 | config=config, depth=depth, date_range=(from_date, to_date), |
| 2573 | runtime=runtime, mock=mock, |
| 2574 | rate_limited_sources=rate_limited_sources, |
| 2575 | rate_limit_lock=rate_limit_lock, |
| 2576 | web_backend=web_backend, |
| 2577 | raw_topic=topic, |
| 2578 | subreddits=subreddits, |
| 2579 | tiktok_hashtags=tiktok_hashtags, |
| 2580 | tiktok_creators=tiktok_creators, |
| 2581 | ig_creators=ig_creators, |
| 2582 | trustpilot_domain=trustpilot_domain, |
| 2583 | trustpilot_domain_is_hint=trustpilot_domain_is_hint, |
| 2584 | run_started=run_started, |
| 2585 | ) |
| 2586 | except Exception as retry_exc: |
| 2587 | detail = f"{exc} (retried once, still failed: {retry_exc})" |
| 2588 | bundle.errors_by_source[source] = detail |
| 2589 | state, attempted = _classify_source_failure(retry_exc) |
| 2590 | bundle.record_failure(source, state, detail, attempted=attempted) |
| 2591 | continue |
| 2592 | else: |
| 2593 | bundle.errors_by_source[source] = str(exc) |
| 2594 | state, attempted = _classify_source_failure(exc) |
| 2595 | bundle.record_failure(source, state, str(exc), attempted=attempted) |
| 2596 | continue |
| 2597 | outcome_note = None |
| 2598 | if isinstance(artifact, dict) and artifact.get("_source_outcome"): |
| 2599 | artifact = dict(artifact) |
| 2600 | outcome_note = artifact.pop("_source_outcome") |
| 2601 | bundle.record_failure( |
| 2602 | source, |
| 2603 | outcome_note["state"], |
| 2604 | outcome_note["detail"], |
| 2605 | attempted=outcome_note.get("attempted", True), |
| 2606 | ) |
| 2607 | if isinstance(artifact, dict) and artifact.get("_source_outcome_detail"): |
| 2608 | artifact = dict(artifact) |
| 2609 | lane_state = artifact.pop("_source_outcome_detail_state", None) |
| 2610 | bundle.record_detail( |
| 2611 | source, artifact.pop("_source_outcome_detail"), state=lane_state |
| 2612 | ) |
| 2613 | if lane_state == health.RATE_LIMITED: |
| 2614 | # Do not re-fan-out against a host still inside its window. |
| 2615 | with rate_limit_lock: |
| 2616 | rate_limited_sources.add(source) |
| 2617 | normalized = _normalize_score_dedupe( |
| 2618 | source, raw_items, from_date, to_date, |
| 2619 | freshness_mode=plan.freshness_mode, |
| 2620 | ranking_query=subquery.ranking_query, |
| 2621 | first_party_handles=explicit_first_party, |
| 2622 | first_party_by_source=creator_first_party, |
| 2623 | # X defers its relevance floor until resolved_handles exists. |
| 2624 | # Everything else prunes here as before. |
| 2625 | defer_relevance_prune=(source == "x"), |
| 2626 | ) |
| 2627 | # Jobs is exempt from per_stream_limit: a careers board is a complete |
| 2628 | # snapshot of open roles, and truncating it to the default 12 drops |
| 2629 | # strategic postings (the whole point of hiring-signals coverage). |
| 2630 | if source != "jobs": |
| 2631 | normalized = _apply_reddit_stream_keepers( |
| 2632 | source, normalized, settings["per_stream_limit"], topic |
| 2633 | ) |
| 2634 | bundle.add_items(subquery.label, source, normalized) |
| 2635 | if artifact: |
| 2636 | bundle.artifacts.setdefault("grounding", []).append(artifact) |
| 2637 | |
| 2638 | # Phase 2: supplemental entity-based searches |
| 2639 | supplemental_handles: list[str] = [] |
| 2640 | _run_supplemental_searches( |
| 2641 | topic=topic, |
| 2642 | bundle=bundle, |
| 2643 | plan=plan, |
| 2644 | config=config, |
| 2645 | depth=depth, |
| 2646 | date_range=(from_date, to_date), |
| 2647 | runtime=runtime, |
| 2648 | mock=mock, |
| 2649 | rate_limited_sources=rate_limited_sources, |
| 2650 | rate_limit_lock=rate_limit_lock, |
| 2651 | x_handle=x_handle, |
| 2652 | x_related=x_related, |
| 2653 | resolved_handles_out=supplemental_handles, |
| 2654 | ) |
| 2655 | |
| 2656 | # Phase 2b: retry thin sources with simplified query |
| 2657 | # Note: _github_skip_sources tells the retry to not re-run GitHub keyword search |
| 2658 | # when project-mode or person-mode already provided authoritative data. |
| 2659 | _github_skip_retry = {"corpus"} |
| 2660 | if _github_person_done or _github_custom_done: |
| 2661 | _github_skip_retry.add("github") |
| 2662 | _retry_thin_sources( |
| 2663 | topic=topic, |
| 2664 | bundle=bundle, |
| 2665 | plan=plan, |
| 2666 | config=config, |
| 2667 | depth=depth, |
| 2668 | date_range=(from_date, to_date), |
| 2669 | runtime=runtime, |
| 2670 | mock=mock, |
| 2671 | rate_limited_sources=rate_limited_sources, |
| 2672 | rate_limit_lock=rate_limit_lock, |
| 2673 | settings=settings, |
| 2674 | web_backend=web_backend, |
| 2675 | skip_sources=_github_skip_retry, |
| 2676 | subreddits=subreddits, |
| 2677 | tiktok_hashtags=tiktok_hashtags, |
| 2678 | tiktok_creators=tiktok_creators, |
| 2679 | ig_creators=ig_creators, |
| 2680 | first_party_handles=explicit_first_party, |
| 2681 | first_party_by_source=creator_first_party, |
| 2682 | run_started=run_started, |
| 2683 | ) |
| 2684 | |
| 2685 | # Reclassify partial failures as DEGRADED instead of silently dropping them. |
| 2686 | # A source that 429'd on one subquery but succeeded on another is not a hard |
| 2687 | # failure, but it is not healthy either: it likely returned fewer results |
| 2688 | # than it should have. Move it out of errors_by_source (so it isn't reported |
| 2689 | # as "failed") and into degraded_by_source (so it survives into warnings), |
| 2690 | # rather than deleting the signal outright as the engine used to. |
| 2691 | degraded_by_source: dict[str, str] = {} |
| 2692 | for source in list(bundle.errors_by_source): |
| 2693 | if bundle.items_by_source.get(source): |
| 2694 | degraded_by_source[source] = bundle.errors_by_source[source] |
| 2695 | del bundle.errors_by_source[source] |
| 2696 | |
| 2697 | hiring_summary = _apply_hiring_signal_gate( |
| 2698 | bundle, |
| 2699 | explicit=hiring_signals_mode, |
| 2700 | topic=topic, |
| 2701 | ) |
| 2702 | if hiring_summary: |
| 2703 | bundle.artifacts["hiring_signals"] = hiring_summary |
| 2704 | |
| 2705 | items_by_source = _finalize_items_by_source( |
| 2706 | bundle.items_by_source, topic=topic, config=config, depth=depth, mock=mock, |
| 2707 | elapsed=time.monotonic() - run_started, |
| 2708 | ) |
| 2709 | source_status = _finalize_source_status(bundle.source_status, items_by_source) |
| 2710 | # Normalized set of handles this run resolved for the topic. A candidate |
| 2711 | # authored by one of these is first-party and is exempted from the |
| 2712 | # entity-miss demotion in rerank (a post never repeats its own author's |
| 2713 | # name, so the body-text grounding check would otherwise zero out the |
| 2714 | # subject's own highest-signal posts). Built before fusion so the |
| 2715 | # per-author cap can give the topic's subject a higher allowance than an |
| 2716 | # incidental third-party account. |
| 2717 | resolved_handles = explicit_first_party | { |
| 2718 | h.lstrip("@").strip().lower() |
| 2719 | for h in supplemental_handles |
| 2720 | if h and h.strip() |
| 2721 | } | {h for handles in creator_first_party.values() for h in handles} |
| 2722 | # resolved_handles feeds rerank/fusion, where the first-party marks only |
| 2723 | # resist demotion of items that already passed the inclusion floors and a |
| 2724 | # named account is treated as the subject across surfaces. The inclusion |
| 2725 | # gate (prune_low_relevance) instead gets the platform-scoped |
| 2726 | # creator_first_party map so a cross-platform name collision cannot |
| 2727 | # bypass the floors. |
| 2728 | # Real X handles from explicit flags, @mentions in topic, or Phase 2 discovery. |
| 2729 | # When no real handle is identified, skip the X floor entirely — a noisier |
| 2730 | # report beats losing the subject's evidence. Topic tokens like "peter" are |
| 2731 | # NOT real handles: they populate resolved_handles for downstream first-party |
| 2732 | # protection but should NOT trigger the floor. |
| 2733 | real_x_handles = explicit_x_handles | { |
| 2734 | h.lstrip("@").strip().lower() |
| 2735 | for h in supplemental_handles |
| 2736 | if h and h.strip() |
| 2737 | } |
| 2738 | # Deferred X relevance floor. Phase 1 skipped it so this could run with the |
| 2739 | # run's actual resolved handles rather than a guess made before anyone knew |
| 2740 | # who the subject was. Applied per subquery stream so fusion sees the same |
| 2741 | # shape it always has. Only applied when we have real X handles — topic |
| 2742 | # tokens alone cannot identify the subject. |
| 2743 | if real_x_handles: |
| 2744 | # IG/TikTok creator exemptions stay on their own platforms: a creator |
| 2745 | # handle must not exempt a same-name X account from this floor. Only |
| 2746 | # creator-ONLY handles are subtracted, where X provenance means |
| 2747 | # real_x_handles (explicit X flags, @mentions in the topic, Phase 2 |
| 2748 | # discovery) - a plain topic token or --github-user match is NOT X |
| 2749 | # provenance and does not preserve the exemption. |
| 2750 | x_floor_handles = resolved_handles - _creator_only_handles( |
| 2751 | creator_first_party, real_x_handles |
| 2752 | ) |
| 2753 | for key, stream in list(bundle.items_by_source_and_query.items()): |
| 2754 | if key[1] != "x" or not stream: |
| 2755 | continue |
| 2756 | pruned = signals.prune_low_relevance( |
| 2757 | stream, |
| 2758 | first_party_handles=x_floor_handles, |
| 2759 | first_party_by_source=creator_first_party, |
| 2760 | ) |
| 2761 | _log_prune_drop("x", len(stream), len(pruned), scope="per-query stream") |
| 2762 | bundle.items_by_source_and_query[key] = pruned |
| 2763 | if bundle.items_by_source.get("x"): |
| 2764 | x_stream = bundle.items_by_source["x"] |
| 2765 | pruned = signals.prune_low_relevance( |
| 2766 | x_stream, |
| 2767 | first_party_handles=x_floor_handles, |
| 2768 | first_party_by_source=creator_first_party, |
| 2769 | ) |
| 2770 | _log_prune_drop("x", len(x_stream), len(pruned), scope="merged stream") |
| 2771 | bundle.items_by_source["x"] = pruned |
| 2772 | |
| 2773 | candidates = weighted_rrf( |
| 2774 | bundle.items_by_source_and_query, |
| 2775 | plan, |
| 2776 | pool_limit=settings["pool_limit"], |
| 2777 | range_from=from_date, |
| 2778 | range_to=to_date, |
| 2779 | first_party_handles=resolved_handles, |
| 2780 | ) |
| 2781 | private_candidates = [ |
| 2782 | candidate |
| 2783 | for candidate in candidates |
| 2784 | if candidate.source == "corpus" |
| 2785 | or any(item.source == "corpus" for item in candidate.source_items) |
| 2786 | ] |
| 2787 | private_candidate_ids = {id(candidate) for candidate in private_candidates} |
| 2788 | public_candidates = [ |
| 2789 | candidate for candidate in candidates if id(candidate) not in private_candidate_ids |
| 2790 | ] |
| 2791 | ranked_public = rerank.rerank_candidates( |
| 2792 | topic=topic, |
| 2793 | plan=plan, |
| 2794 | candidates=public_candidates, |
| 2795 | provider=None if mock else reasoning_provider, |
| 2796 | model=None if mock else runtime.rerank_model, |
| 2797 | shortlist_size=settings["rerank_limit"], |
| 2798 | resolved_handles=resolved_handles, |
| 2799 | ) |
| 2800 | # Corpus titles/snippets must never enter a hosted reasoning prompt. Score |
| 2801 | # every candidate carrying corpus evidence with the deterministic fallback, |
| 2802 | # even when the rest of the run uses a remote reranker. |
| 2803 | ranked_private = rerank.rerank_candidates( |
| 2804 | topic=topic, |
| 2805 | plan=plan, |
| 2806 | candidates=private_candidates, |
| 2807 | provider=None, |
| 2808 | model=None, |
| 2809 | shortlist_size=settings["rerank_limit"], |
| 2810 | resolved_handles=resolved_handles, |
| 2811 | ) |
| 2812 | ranked_public = rerank.prune_fallback_entity_misses(ranked_public, topic=topic) |
| 2813 | if hiring_summary: |
| 2814 | # The diagnostic source dump retains pruned jobs; hiring aggregation |
| 2815 | # needs their rejection identities after they leave the ranked pool. |
| 2816 | retained_ids = {candidate.candidate_id for candidate in ranked_public} |
| 2817 | hiring_summary["rejected_job_keys"] = sorted({ |
| 2818 | fusion.candidate_key(item) |
| 2819 | for candidate in public_candidates |
| 2820 | if candidate.candidate_id not in retained_ids |
| 2821 | for item in candidate.source_items |
| 2822 | if item.source == "jobs" |
| 2823 | }) |
| 2824 | # Private corpus already cleared a body-aware retrieval floor; do not apply |
| 2825 | # the public title/snippet visibility gate (filenames often omit the head |
| 2826 | # token even when the document body matched). |
| 2827 | ranked_candidates = sorted( |
| 2828 | [*ranked_public, *ranked_private], |
| 2829 | key=lambda candidate: ( |
| 2830 | 1 if schema.candidate_out_of_window(candidate) else 0, |
| 2831 | -candidate.final_score, |
| 2832 | -(candidate.engagement or -1), |
| 2833 | min(candidate.native_ranks.values(), default=999), |
| 2834 | candidate.title, |
| 2835 | ), |
| 2836 | ) |
| 2837 | rerank.score_fun( |
| 2838 | topic=topic, |
| 2839 | candidates=ranked_public, |
| 2840 | provider=None if mock else reasoning_provider, |
| 2841 | model=None if mock else runtime.rerank_model, |
| 2842 | ) |
| 2843 | rerank.score_fun( |
| 2844 | topic=topic, |
| 2845 | candidates=ranked_private, |
| 2846 | provider=None, |
| 2847 | model=None, |
| 2848 | ) |
| 2849 | |
| 2850 | # Phase 3: post-rerank GitHub star enrichment. Record/replay-aware so the |
| 2851 | # eval harness stays fully offline: this path calls the GitHub API (and the |
| 2852 | # gh-credential fallback) outside the _retrieve_stream seam, so it gets its |
| 2853 | # own fixture exchange keyed by phase. |
| 2854 | if "github" in available and not mock: |
| 2855 | star_request = { |
| 2856 | "source": "github", |
| 2857 | "phase": "post_rerank_star_enrichment", |
| 2858 | "topic": topic, |
| 2859 | "depth": depth, |
| 2860 | } |
| 2861 | star_matched, star_replayed = http.fixture_source_replay(star_request) |
| 2862 | if star_matched: |
| 2863 | star_map = star_replayed if isinstance(star_replayed, dict) else {} |
| 2864 | github.apply_star_map(ranked_candidates, star_map) |
| 2865 | else: |
| 2866 | collected_star_map: dict[str, int] = {} |
| 2867 | github.enrich_candidates_with_stars( |
| 2868 | ranked_candidates, |
| 2869 | token=config.get("GITHUB_TOKEN"), |
| 2870 | already_enriched=_github_enriched_repos, |
| 2871 | collect_map=collected_star_map, |
| 2872 | ) |
| 2873 | http.fixture_source_record(star_request, collected_star_map) |
| 2874 | |
| 2875 | clusters = cluster_candidates(ranked_candidates, plan) |
| 2876 | warnings = _warnings(items_by_source, ranked_candidates, bundle.errors_by_source, degraded_by_source) |
| 2877 | # One-sided entity coverage is a reporting warning, not a source failure: |
| 2878 | # marking the source PARTIAL would trip LAST30DAYS_STRICT_EXIT on runs that |
| 2879 | # returned good X results. |
| 2880 | warnings.extend(bundle.artifacts.get("x_partial_coverage", [])) |
| 2881 | # Backend receipts that are not failures (xapi's truncated window), and |
| 2882 | # the Meta Ads footer inputs. A stream artifact only ever reaches the |
| 2883 | # report as an anonymous entry in this list, so the advertiser and the |
| 2884 | # pre-truncation counts have to be lifted to named top-level artifacts or |
| 2885 | # the footer cannot render them -- least of all on a zero-item run, which |
| 2886 | # is exactly when naming the advertiser matters most. |
| 2887 | for stream_artifact in bundle.artifacts.get("grounding", []): |
| 2888 | if isinstance(stream_artifact, dict): |
| 2889 | warnings.extend(stream_artifact.get("x_receipts", [])) |
| 2890 | _lift_stream_artifacts(bundle) |
| 2891 | library_context, library_warning = _load_library_context( |
| 2892 | topic=topic, |
| 2893 | config=config, |
| 2894 | mock=mock, |
| 2895 | internal_subrun=internal_subrun, |
| 2896 | x_handle=x_handle, |
| 2897 | github_user=github_user, |
| 2898 | github_repos=github_repos, |
| 2899 | save_dir=save_dir, |
| 2900 | ) |
| 2901 | if library_warning: |
| 2902 | warnings.append(library_warning) |
| 2903 | |
| 2904 | return schema.Report( |
| 2905 | topic=topic, |
| 2906 | range_from=from_date, |
| 2907 | range_to=to_date, |
| 2908 | generated_at=datetime.now(timezone.utc).isoformat(), |
| 2909 | provider_runtime=runtime, |
| 2910 | query_plan=plan, |
| 2911 | clusters=clusters, |
| 2912 | ranked_candidates=ranked_candidates, |
| 2913 | items_by_source=items_by_source, |
| 2914 | errors_by_source=bundle.errors_by_source, |
| 2915 | source_status=source_status, |
| 2916 | warnings=warnings, |
| 2917 | artifacts=bundle.artifacts, |
| 2918 | library_context=library_context, |
| 2919 | ) |
| 2920 | |
| 2921 | |
| 2922 | def _candidate_is_duplicate( |
| 2923 | candidate: schema.Candidate, |
| 2924 | kept: list[schema.Candidate], |
| 2925 | ) -> bool: |
| 2926 | if any(existing.candidate_id == candidate.candidate_id for existing in kept): |
| 2927 | return True |
| 2928 | if candidate.url and any(existing.url == candidate.url for existing in kept): |
| 2929 | return True |
| 2930 | candidate_text = " ".join((candidate.title, candidate.snippet)).strip() |
| 2931 | return bool(candidate_text) and any( |
| 2932 | dedupe.hybrid_similarity( |
| 2933 | candidate_text, |
| 2934 | " ".join((existing.title, existing.snippet)).strip(), |
| 2935 | ) >= 0.7 |
| 2936 | for existing in kept |
| 2937 | ) |
| 2938 | |
| 2939 | |
| 2940 | def merge_drill_report( |
| 2941 | report: schema.Report, |
| 2942 | drill_report: schema.Report, |
| 2943 | matched_clusters: list[schema.Cluster], |
| 2944 | *, |
| 2945 | target: str, |
| 2946 | ) -> schema.Report: |
| 2947 | """Merge a narrow follow-up into its cached report while preserving other clusters.""" |
| 2948 | merged = copy.deepcopy(report) |
| 2949 | selected_cluster_ids = {cluster.cluster_id for cluster in matched_clusters} |
| 2950 | selected_candidate_ids = { |
| 2951 | candidate_id |
| 2952 | for cluster in matched_clusters |
| 2953 | for candidate_id in cluster.candidate_ids |
| 2954 | } |
| 2955 | original_candidates = { |
| 2956 | candidate.candidate_id: candidate for candidate in merged.ranked_candidates |
| 2957 | } |
| 2958 | unrelated_candidates = [ |
| 2959 | candidate for candidate in merged.ranked_candidates |
| 2960 | if candidate.candidate_id not in selected_candidate_ids |
| 2961 | ] |
| 2962 | original_summary = "" |
| 2963 | for cluster in matched_clusters: |
| 2964 | for candidate_id in cluster.representative_ids: |
| 2965 | candidate = original_candidates.get(candidate_id) |
| 2966 | if candidate: |
| 2967 | original_summary = candidate.snippet or candidate.explanation or candidate.title |
| 2968 | if original_summary: |
| 2969 | break |
| 2970 | if original_summary: |
| 2971 | break |
| 2972 | |
| 2973 | unrelated_candidate_indexes = { |
| 2974 | candidate.candidate_id: index |
| 2975 | for index, candidate in enumerate(unrelated_candidates) |
| 2976 | } |
| 2977 | focused_candidates: list[schema.Candidate] = [] |
| 2978 | for candidate in [ |
| 2979 | *copy.deepcopy(drill_report.ranked_candidates), |
| 2980 | *[ |
| 2981 | copy.deepcopy(candidate) |
| 2982 | for candidate in merged.ranked_candidates |
| 2983 | if candidate.candidate_id in selected_candidate_ids |
| 2984 | ], |
| 2985 | ]: |
| 2986 | unrelated_index = unrelated_candidate_indexes.get(candidate.candidate_id) |
| 2987 | if unrelated_index is not None: |
| 2988 | candidate.cluster_id = unrelated_candidates[unrelated_index].cluster_id |
| 2989 | unrelated_candidates[unrelated_index] = candidate |
| 2990 | continue |
| 2991 | if not _candidate_is_duplicate(candidate, focused_candidates): |
| 2992 | focused_candidates.append(candidate) |
| 2993 | |
| 2994 | primary_cluster = matched_clusters[0] |
| 2995 | for candidate in focused_candidates: |
| 2996 | candidate.cluster_id = primary_cluster.cluster_id |
| 2997 | focused_ids = [candidate.candidate_id for candidate in focused_candidates] |
| 2998 | focused_sources = sorted({ |
| 2999 | source |
| 3000 | for candidate in focused_candidates |
| 3001 | for source in schema.candidate_sources(candidate) |
| 3002 | }) |
| 3003 | replacement_cluster = schema.Cluster( |
| 3004 | cluster_id=primary_cluster.cluster_id, |
| 3005 | title=primary_cluster.title, |
| 3006 | candidate_ids=focused_ids, |
| 3007 | representative_ids=focused_ids[:3], |
| 3008 | sources=focused_sources, |
| 3009 | score=max((candidate.final_score for candidate in focused_candidates), default=0.0), |
| 3010 | uncertainty="single-source" if len(focused_sources) == 1 else None, |
| 3011 | ) |
| 3012 | |
| 3013 | first_selected_index = min( |
| 3014 | index |
| 3015 | for index, cluster in enumerate(merged.clusters) |
| 3016 | if cluster.cluster_id in selected_cluster_ids |
| 3017 | ) |
| 3018 | remaining_clusters = [ |
| 3019 | cluster for cluster in merged.clusters |
| 3020 | if cluster.cluster_id not in selected_cluster_ids |
| 3021 | ] |
| 3022 | remaining_clusters.insert(first_selected_index, replacement_cluster) |
| 3023 | merged.clusters = remaining_clusters |
| 3024 | |
| 3025 | merged.ranked_candidates = focused_candidates + unrelated_candidates |
| 3026 | |
| 3027 | all_sources = set(merged.items_by_source) | set(drill_report.items_by_source) |
| 3028 | new_item_count = 0 |
| 3029 | merged_items: dict[str, list[schema.SourceItem]] = {} |
| 3030 | for source in sorted(all_sources): |
| 3031 | old_items = merged.items_by_source.get(source, []) |
| 3032 | new_items = drill_report.items_by_source.get(source, []) |
| 3033 | # Collapse exact URL matches first, preferring the drill's copy (it |
| 3034 | # carries fresh transcripts/comments); fuzzy dedupe alone keeps both |
| 3035 | # when enrichment changed the text substantially. |
| 3036 | new_urls = {item.url for item in new_items if item.url} |
| 3037 | kept_old = [item for item in old_items if not (item.url and item.url in new_urls)] |
| 3038 | combined = dedupe.dedupe_items([*copy.deepcopy(new_items), *kept_old]) |
| 3039 | old_unique = dedupe.dedupe_items(old_items) |
| 3040 | new_item_count += max(0, len(combined) - len(old_unique)) |
| 3041 | merged_items[source] = combined |
| 3042 | merged.items_by_source = merged_items |
| 3043 | |
| 3044 | merged.generated_at = drill_report.generated_at |
| 3045 | merged.query_plan = drill_report.query_plan |
| 3046 | # The drill's retrieval window is the report's window now (a --days/--as-of |
| 3047 | # override on the drill must not be mislabeled with the cached range). |
| 3048 | merged.range_from = drill_report.range_from |
| 3049 | merged.range_to = drill_report.range_to |
| 3050 | attempted_sources = { |
| 3051 | source |
| 3052 | for source, outcome in drill_report.source_status.items() |
| 3053 | if outcome.attempted or outcome.state == schema.SKIPPED_UNCONFIGURED |
| 3054 | } |
| 3055 | for source in attempted_sources: |
| 3056 | if source in drill_report.errors_by_source: |
| 3057 | merged.errors_by_source[source] = drill_report.errors_by_source[source] |
| 3058 | else: |
| 3059 | merged.errors_by_source.pop(source, None) |
| 3060 | merged.source_status[source] = drill_report.source_status[source] |
| 3061 | merged.source_status = _finalize_source_status( |
| 3062 | merged.source_status, |
| 3063 | merged.items_by_source, |
| 3064 | ) |
| 3065 | degraded_by_source = { |
| 3066 | source: outcome.detail or "partial results" |
| 3067 | for source, outcome in merged.source_status.items() |
| 3068 | if outcome.state == schema.PARTIAL |
| 3069 | } |
| 3070 | merged.warnings = _warnings( |
| 3071 | merged.items_by_source, |
| 3072 | merged.ranked_candidates, |
| 3073 | merged.errors_by_source, |
| 3074 | degraded_by_source, |
| 3075 | ) |
| 3076 | merged.artifacts.update(copy.deepcopy(drill_report.artifacts)) |
| 3077 | history = list(merged.artifacts.get("drill_history") or []) |
| 3078 | history.append({ |
| 3079 | "target": target, |
| 3080 | "clusters": [cluster.title for cluster in matched_clusters], |
| 3081 | "new_items": new_item_count, |
| 3082 | "generated_at": drill_report.generated_at, |
| 3083 | }) |
| 3084 | merged.artifacts["drill_history"] = history |
| 3085 | merged.artifacts["drill_context"] = { |
| 3086 | "target": target, |
| 3087 | "cluster_titles": [cluster.title for cluster in matched_clusters], |
| 3088 | "original_summary": original_summary, |
| 3089 | "new_items": new_item_count, |
| 3090 | "sources": focused_sources, |
| 3091 | } |
| 3092 | merged.drill_of = primary_cluster.title |
| 3093 | return merged |
| 3094 | |
| 3095 | |
| 3096 | def _batch_subject_handles(raw_items: list[dict], *, top_n: int = 2) -> set[str]: |
| 3097 | """Most-mentioned handles in a batch of X items, as first-party candidates. |
| 3098 | |
| 3099 | Mirrors entity_extract's ranking but runs before pruning rather than after, |
| 3100 | and keys on *mentions only* rather than mentions plus authors. That |
| 3101 | distinction is the safety property: a prolific commentator inflates the |
| 3102 | author count, but being mentioned by other accounts is what identifies the |
| 3103 | subject of a topic. Capped at the top few so a busy thread cannot exempt |
| 3104 | the whole batch. |
| 3105 | """ |
| 3106 | counts: Counter = Counter() |
| 3107 | for item in raw_items or []: |
| 3108 | text = str((item or {}).get("text") or "") |
| 3109 | for mention in re.findall(r"@([A-Za-z0-9_]{1,15})", text): |
| 3110 | counts[mention.lower()] += 1 |
| 3111 | if not counts: |
| 3112 | return set() |
| 3113 | return {handle for handle, _ in counts.most_common(top_n)} |
| 3114 | |
| 3115 | |
| 3116 | # Reddit engagement keepers: per stream, the top-N threads by upvotes plus |
| 3117 | # comments that clear the relevance floor and name the primary entity survive |
| 3118 | # per_stream_limit truncation even when their local rank score is low. The |
| 3119 | # stream order is 65% title relevance, so the month's most-discussed on-topic |
| 3120 | # thread (16K upvotes, 0.19 relevance) was otherwise cut behind one-upvote |
| 3121 | # posts with better title overlap. |
| 3122 | REDDIT_STREAM_KEEPERS = 3 |
| 3123 | |
| 3124 | |
| 3125 | def _apply_reddit_stream_keepers( |
| 3126 | source: str, |
| 3127 | items: list[schema.SourceItem], |
| 3128 | limit: int, |
| 3129 | topic: str, |
| 3130 | ) -> list[schema.SourceItem]: |
| 3131 | """Truncate a stream to *limit*, holding slots for Reddit engagement keepers.""" |
| 3132 | kept = list(items[:limit]) |
| 3133 | if source != "reddit" or len(items) <= limit: |
| 3134 | return kept |
| 3135 | entity = rerank._primary_entity(topic or "") if topic else "" |
| 3136 | floor = fusion.relevance_floor_for_entity(entity) |
| 3137 | keepers = [ |
| 3138 | item |
| 3139 | for item in sorted(items, key=fusion.raw_engagement, reverse=True) |
| 3140 | if fusion.reddit_thread_qualifies(item, entity, floor) |
| 3141 | ][:REDDIT_STREAM_KEEPERS] |
| 3142 | keeper_ids = {id(item) for item in keepers} |
| 3143 | for keeper in keepers: |
| 3144 | if any(item is keeper for item in kept): |
| 3145 | continue |
| 3146 | # Displace the lowest-ranked non-keeper so the slice stays at limit; |
| 3147 | # when the slice is already all keepers there is nothing to trade. |
| 3148 | displaced = False |
| 3149 | for index in range(len(kept) - 1, -1, -1): |
| 3150 | if id(kept[index]) not in keeper_ids: |
| 3151 | del kept[index] |
| 3152 | displaced = True |
| 3153 | break |
| 3154 | if displaced or len(kept) < limit: |
| 3155 | kept.append(keeper) |
| 3156 | return kept[:limit] |
| 3157 | |
| 3158 | |
| 3159 | |
| 3160 | def _creator_first_party_by_source( |
| 3161 | tiktok_creators: Iterable[str] | None, |
| 3162 | ig_creators: Iterable[str] | None, |
| 3163 | ) -> dict[str, set[str]]: |
| 3164 | """Platform-scoped creator handles for the relevance prune. |
| 3165 | |
| 3166 | --ig-creators names Instagram accounts and --creators names TikTok |
| 3167 | accounts; each exemption applies only on its own platform. Merging both |
| 3168 | into one global handle set would let an unrelated same-name account on |
| 3169 | another platform bypass the relevance and engagement floors. |
| 3170 | """ |
| 3171 | |
| 3172 | def _norm(handles: Iterable[str] | None) -> set[str]: |
| 3173 | return { |
| 3174 | h.lstrip("@").strip().lower() |
| 3175 | for h in (handles or []) |
| 3176 | if h and h.strip() |
| 3177 | } |
| 3178 | |
| 3179 | return {"instagram": _norm(ig_creators), "tiktok": _norm(tiktok_creators)} |
| 3180 | |
| 3181 | |
| 3182 | def _creator_only_handles( |
| 3183 | creator_first_party: Mapping[str, set[str]], |
| 3184 | *x_provenance_sets: Iterable[str], |
| 3185 | ) -> set[str]: |
| 3186 | """Creator handles that carry no X provenance. |
| 3187 | |
| 3188 | A handle named ONLY via --ig-creators / --creators must not exempt a |
| 3189 | same-name X account from the deferred X floor. But the same person is |
| 3190 | often named on both surfaces (--x-handle foo --ig-creators foo): the |
| 3191 | normalized sets collapse that to one string, so subtracting the whole |
| 3192 | creator set would strip the explicitly requested X exemption too. The |
| 3193 | subtraction therefore covers only handles absent from every |
| 3194 | X-provenance set (explicit flags, topic mentions, Phase 2 discovery). |
| 3195 | """ |
| 3196 | creator_flat = {h for handles in creator_first_party.values() for h in handles} |
| 3197 | x_provenance = {h for handles in x_provenance_sets for h in handles} |
| 3198 | return creator_flat - x_provenance |
| 3199 | |
| 3200 | |
| 3201 | def _log_prune_drop( |
| 3202 | source: str, before: int, after: int, scope: str | None = None |
| 3203 | ) -> None: |
| 3204 | """Log when the relevance prune removes items from a stream. |
| 3205 | |
| 3206 | The prune is silent by design inside ``signals`` (a pure function), but a |
| 3207 | silent drop is invisible to the user: issue #1101 fetched 36 creator reels |
| 3208 | and reported zero with no line explaining why. Log the count and the reason |
| 3209 | class here, next to the other per-stream retrieval logs. The all-weak |
| 3210 | ``filtered or items`` rescue keeps the originals, so before == after and |
| 3211 | nothing is logged - a rescue is not a drop. |
| 3212 | """ |
| 3213 | dropped = before - after |
| 3214 | if dropped <= 0: |
| 3215 | return |
| 3216 | scope_note = f" ({scope})" if scope else "" |
| 3217 | log.source_log( |
| 3218 | render.SOURCE_LABELS.get(source, source.capitalize()), |
| 3219 | f"relevance prune dropped {dropped} of {before} items below the " |
| 3220 | f"relevance/engagement floor{scope_note}", |
| 3221 | tty_only=False, |
| 3222 | ) |
| 3223 | |
| 3224 | |
| 3225 | def _normalize_score_dedupe( |
| 3226 | source: str, |
| 3227 | raw_items: list[dict], |
| 3228 | from_date: str, |
| 3229 | to_date: str, |
| 3230 | freshness_mode: str, |
| 3231 | ranking_query: str, |
| 3232 | first_party_handles: Iterable[str] | None = None, |
| 3233 | first_party_by_source: Mapping[str, Iterable[str]] | None = None, |
| 3234 | defer_relevance_prune: bool = False, |
| 3235 | ) -> list[schema.SourceItem]: |
| 3236 | """Normalize, annotate, prune, dedupe, and extract snippets for a batch of raw items. |
| 3237 | |
| 3238 | ``defer_relevance_prune`` skips the relevance floor here so the caller can |
| 3239 | apply it once the run has resolved who the topic's subject is. Pruning X |
| 3240 | before handle resolution is the ordering bug behind the whole first-party |
| 3241 | evidence loss: the floor cannot exempt an author nobody has identified yet, |
| 3242 | and no amount of guessing at prune time substitutes for knowing. |
| 3243 | |
| 3244 | ``first_party_handles`` names accounts this run is explicitly searching, so |
| 3245 | their own posts survive the relevance floor (see signals.prune_low_relevance). |
| 3246 | """ |
| 3247 | normalized = normalize.normalize_source_items( |
| 3248 | source, raw_items, from_date, to_date, |
| 3249 | freshness_mode=freshness_mode, |
| 3250 | ) |
| 3251 | prepared_query = relevance.PreparedQuery(ranking_query) |
| 3252 | lookback_window_days = ( |
| 3253 | datetime.strptime(to_date, "%Y-%m-%d").date() |
| 3254 | - datetime.strptime(from_date, "%Y-%m-%d").date() |
| 3255 | ).days |
| 3256 | normalized = signals.annotate_stream( |
| 3257 | normalized, |
| 3258 | prepared_query, |
| 3259 | freshness_mode, |
| 3260 | reference_date=to_date, |
| 3261 | max_days=lookback_window_days, |
| 3262 | ) |
| 3263 | if source != "jobs" and not defer_relevance_prune: |
| 3264 | floor_handles = set(first_party_handles or ()) |
| 3265 | if source == "x": |
| 3266 | # Union, never a fallback. The caller's set is derived partly from |
| 3267 | # topic tokens, so it is non-empty for essentially every real topic |
| 3268 | # -- gating this on "no handles supplied" would make it dead code |
| 3269 | # and leave the name-only case exactly as broken as before. |
| 3270 | # |
| 3271 | # Reuses the engine's own resolution signal on the batch already in |
| 3272 | # hand: posts *about* a subject mention their handle, so the |
| 3273 | # most-mentioned account in a topic's own results is the subject. |
| 3274 | # Costs nothing extra -- no search, no network -- and closes the |
| 3275 | # case where the handle never appears in the topic at all |
| 3276 | # ("Peter Steinberger" -> @steipete). |
| 3277 | floor_handles |= _batch_subject_handles(raw_items) |
| 3278 | pre_prune_count = len(normalized) |
| 3279 | normalized = signals.prune_low_relevance( |
| 3280 | normalized, |
| 3281 | first_party_handles=floor_handles, |
| 3282 | first_party_by_source=first_party_by_source, |
| 3283 | ) |
| 3284 | _log_prune_drop(source, pre_prune_count, len(normalized)) |
| 3285 | normalized = dedupe.dedupe_items(normalized) |
| 3286 | for item in normalized: |
| 3287 | item.snippet = snippet.extract_best_snippet(item, prepared_query) |
| 3288 | return normalized |
| 3289 | |
| 3290 | |
| 3291 | def _finalize_items_by_source( |
| 3292 | items_by_source_raw: dict[str, list[schema.SourceItem]], |
| 3293 | topic: str = "", |
| 3294 | config: dict | None = None, |
| 3295 | depth: str = "default", |
| 3296 | mock: bool = False, |
| 3297 | elapsed: float = 0.0, |
| 3298 | ) -> dict[str, list[schema.SourceItem]]: |
| 3299 | finalized = {} |
| 3300 | for source, items in items_by_source_raw.items(): |
| 3301 | items = sorted(items, key=lambda item: item.local_rank_score or 0.0, reverse=True) |
| 3302 | # Same thread from two subquery streams: fold the enriched copy into |
| 3303 | # the first before the text-similarity dedupe, which would otherwise |
| 3304 | # keep whichever copy ranked higher and drop its comments. |
| 3305 | items = collapse_duplicate_urls(items) |
| 3306 | items = dedupe.dedupe_items(items) |
| 3307 | enrichment_request = { |
| 3308 | "source": source, |
| 3309 | "phase": "post_ranking_enrichment", |
| 3310 | "topic": topic, |
| 3311 | "depth": depth, |
| 3312 | } |
| 3313 | if source == "youtube" and items and not mock: |
| 3314 | # Same budget-at-the-survivors principle as the digg branch |
| 3315 | # below: retrieval-time transcripts go to each search's |
| 3316 | # top-by-views candidates, while final selection ranks by |
| 3317 | # relevance. Backfill survivors that arrived without one so the |
| 3318 | # transcript budget lands on videos the brief actually shows |
| 3319 | # (#542). |
| 3320 | matched, replayed = http.fixture_source_replay(enrichment_request) |
| 3321 | if matched: |
| 3322 | items = _merge_replayed_enrichment(items, replayed) |
| 3323 | else: |
| 3324 | sc_token = ( |
| 3325 | config.get("SCRAPECREATORS_API_KEY") |
| 3326 | if config and env.is_youtube_sc_available(config) else None |
| 3327 | ) |
| 3328 | youtube_yt.backfill_transcripts( |
| 3329 | items, topic=topic, depth=depth, token=sc_token, |
| 3330 | ) |
| 3331 | http.fixture_source_record(enrichment_request, schema.to_dict(items)) |
| 3332 | # Post-merge topic-relevance filter for Polymarket: comparison queries |
| 3333 | # fan out into per-entity subqueries ("Hermes", "OpenClaw") whose topic |
| 3334 | # is too narrow for Gamma API to filter meaningfully. Re-validating the |
| 3335 | # merged list against the full original topic drops off-topic markets |
| 3336 | # (e.g., WTI crude oil, Elon tweet counts) before footer emission. |
| 3337 | if source == "polymarket" and topic: |
| 3338 | items = polymarket.filter_items_against_topic(topic, items) |
| 3339 | # --polymarket-keywords (via config): additional keyword filter |
| 3340 | # for ambiguous single-token topics (e.g., "Warriors" → nba,gsw). |
| 3341 | keywords = config.get("_polymarket_keywords") if isinstance(config, dict) else None |
| 3342 | if keywords: |
| 3343 | items = polymarket.filter_items_against_keywords(items, keywords) |
| 3344 | if source == "digg" and items: |
| 3345 | # Pull top-ranked X posts only for the survivors that will appear |
| 3346 | # in the brief. Spending the enrichment budget here (rather than |
| 3347 | # at retrieval time) keeps the inline 'via Digg' quotes |
| 3348 | # paired with the clusters dedupe actually kept. |
| 3349 | matched, replayed = http.fixture_source_replay(enrichment_request) |
| 3350 | if matched: |
| 3351 | items = _merge_replayed_enrichment(items, replayed) |
| 3352 | else: |
| 3353 | digg.enrich_source_items(items, top_k=3) |
| 3354 | http.fixture_source_record(enrichment_request, schema.to_dict(items)) |
| 3355 | if source == "amazon" and items and not mock: |
| 3356 | # Attach-if-missing: review enrichment now runs at search time in |
| 3357 | # _retrieve_stream_impl, so items arriving here should already have |
| 3358 | # top_comments. enrich_source_items no-ops when top_comments is set. |
| 3359 | # This path handles fixture replay and any edge cases where retrieve |
| 3360 | # didn't enrich (e.g., run_started was not passed). |
| 3361 | matched, replayed = http.fixture_source_replay(enrichment_request) |
| 3362 | if matched: |
| 3363 | items = _merge_replayed_enrichment(items, replayed) |
| 3364 | else: |
| 3365 | amazon.enrich_source_items( |
| 3366 | items, |
| 3367 | depth=depth, |
| 3368 | config=config, |
| 3369 | keyword=str((config or {}).get("_amazon_query") or "").strip() or topic, |
| 3370 | elapsed=elapsed, |
| 3371 | ) |
| 3372 | http.fixture_source_record(enrichment_request, schema.to_dict(items)) |
| 3373 | finalized[source] = items |
| 3374 | return finalized |
| 3375 | |
| 3376 | |
| 3377 | def _merge_replayed_enrichment( |
| 3378 | items: list[schema.SourceItem], |
| 3379 | replayed: list[dict], |
| 3380 | ) -> list[schema.SourceItem]: |
| 3381 | """Apply recorded post-ranking enrichment onto freshly computed items. |
| 3382 | |
| 3383 | Enrichment (transcripts, Digg posts) only mutates ``metadata``. Merging by |
| 3384 | item_id instead of replacing the list keeps normalization, scoring, and |
| 3385 | dedupe regressions visible to the eval - fixture state must not overwrite |
| 3386 | what the current pipeline computed. |
| 3387 | """ |
| 3388 | replayed_by_id = { |
| 3389 | entry.get("item_id"): entry for entry in replayed if isinstance(entry, dict) |
| 3390 | } |
| 3391 | for item in items: |
| 3392 | record = replayed_by_id.get(item.item_id) |
| 3393 | if record and record.get("metadata"): |
| 3394 | item.metadata.update(record["metadata"]) |
| 3395 | return items |
| 3396 | |
| 3397 | |
| 3398 | def _apply_hiring_signal_gate( |
| 3399 | bundle: schema.RetrievalBundle, |
| 3400 | *, |
| 3401 | explicit: bool, |
| 3402 | topic: str, |
| 3403 | ) -> dict[str, Any] | None: |
| 3404 | jobs_items = bundle.items_by_source.get("jobs") or [] |
| 3405 | if not jobs_items: |
| 3406 | if explicit: |
| 3407 | return hiring_signals.analyze([], explicit=True, topic=topic) |
| 3408 | return None |
| 3409 | |
| 3410 | summary = hiring_signals.analyze(jobs_items, explicit=explicit, topic=topic) |
| 3411 | if not explicit and not summary.get("include"): |
| 3412 | bundle.items_by_source.pop("jobs", None) |
| 3413 | for key in list(bundle.items_by_source_and_query): |
| 3414 | if key[1] == "jobs": |
| 3415 | del bundle.items_by_source_and_query[key] |
| 3416 | return summary |
| 3417 | |
| 3418 | |
| 3419 | def _ensure_jobs_in_plan( |
| 3420 | plan: schema.QueryPlan, |
| 3421 | available: list[str], |
| 3422 | *, |
| 3423 | explicit: bool, |
| 3424 | topic: str, |
| 3425 | ) -> None: |
| 3426 | if "jobs" not in available: |
| 3427 | return |
| 3428 | if not (explicit or _company_topic_likely(topic)): |
| 3429 | return |
| 3430 | if "jobs" not in plan.source_weights: |
| 3431 | plan.source_weights["jobs"] = 1.0 |
| 3432 | for subquery in plan.subqueries: |
| 3433 | if "jobs" not in subquery.sources: |
| 3434 | subquery.sources.append("jobs") |
| 3435 | |
| 3436 | |
| 3437 | def _ensure_perplexity_in_plan( |
| 3438 | plan: schema.QueryPlan, |
| 3439 | topic: str, |
| 3440 | available: list[str], |
| 3441 | *, |
| 3442 | force: bool, |
| 3443 | ) -> None: |
| 3444 | """Route a bounded paid Perplexity action through the whole topic. |
| 3445 | |
| 3446 | Deep Research forces its explicit lane. Normal modes are rerouted only when |
| 3447 | the sanitized plan already selected Perplexity. |
| 3448 | """ |
| 3449 | if "perplexity" not in available: |
| 3450 | return |
| 3451 | planned = any( |
| 3452 | "perplexity" in subquery.sources for subquery in plan.subqueries |
| 3453 | ) |
| 3454 | if not force and not planned: |
| 3455 | return |
| 3456 | retained: list[schema.SubQuery] = [] |
| 3457 | for subquery in plan.subqueries: |
| 3458 | sources = [ |
| 3459 | source for source in subquery.sources if source != "perplexity" |
| 3460 | ] |
| 3461 | if sources: |
| 3462 | retained.append(replace(subquery, sources=sources)) |
| 3463 | retained.append( |
| 3464 | schema.SubQuery( |
| 3465 | label="deep-research" if force else "perplexity-whole-topic", |
| 3466 | search_query=topic, |
| 3467 | ranking_query=f"What current source-grounded evidence matters for {topic}?", |
| 3468 | sources=["perplexity"], |
| 3469 | weight=1.0, |
| 3470 | ), |
| 3471 | ) |
| 3472 | plan.subqueries = planner._normalize_subquery_weights(retained) |
| 3473 | plan.source_weights.setdefault("perplexity", 1.0) |
| 3474 | plan.source_weights = planner._normalize_weights(plan.source_weights) |
| 3475 | |
| 3476 | |
| 3477 | def _company_topic_likely(topic: str) -> bool: |
| 3478 | text = topic.strip() |
| 3479 | if not text: |
| 3480 | return False |
| 3481 | lower = text.lower() |
| 3482 | if "?" in text or len(text.split()) > 4: |
| 3483 | return False |
| 3484 | generic = { |
| 3485 | "how", "what", "why", "best", "top", "tutorial", "guide", "prompts", |
| 3486 | "news", "latest", "ideas", "examples", |
| 3487 | } |
| 3488 | if any(word in generic for word in lower.split()): |
| 3489 | return False |
| 3490 | known_single_word_companies = { |
| 3491 | "apple", "uber", "google", "microsoft", "amazon", "meta", "netflix", |
| 3492 | "openai", "anthropic", "qualtrics", "stripe", "brex", |
| 3493 | } |
| 3494 | if " vs " in lower or " versus " in lower: |
| 3495 | parts = re.split(r"\s+(?:vs|versus)\s+", text, maxsplit=1, flags=re.IGNORECASE) |
| 3496 | if len(parts) != 2: |
| 3497 | return False |
| 3498 | return _comparison_side_company_like(parts[0], known_single_word_companies) or _comparison_side_company_like( |
| 3499 | parts[1], known_single_word_companies |
| 3500 | ) |
| 3501 | return bool(text[:1].isupper() or lower in known_single_word_companies) |
| 3502 | |
| 3503 | |
| 3504 | def _comparison_side_company_like(side: str, known_companies: set[str]) -> bool: |
| 3505 | token = re.sub(r"[^\w.+#-]", "", side.strip().split()[0] if side.strip() else "") |
| 3506 | if not token: |
| 3507 | return False |
| 3508 | lower = token.lower() |
| 3509 | common_tech_terms = { |
| 3510 | "python", "ruby", "javascript", "typescript", "java", "go", "golang", |
| 3511 | "rust", "php", "swift", "kotlin", "scala", "clojure", "elixir", |
| 3512 | "react", "vue", "angular", "svelte", "node", "django", "rails", |
| 3513 | "postgres", "mysql", "redis", "kubernetes", "docker", |
| 3514 | } |
| 3515 | if lower in common_tech_terms: |
| 3516 | return False |
| 3517 | return bool(token[:1].isupper() or lower in known_companies) |
| 3518 | |
| 3519 | |
| 3520 | def _warnings( |
| 3521 | items_by_source: dict[str, list[schema.SourceItem]], |
| 3522 | candidates: list[schema.Candidate], |
| 3523 | errors_by_source: dict[str, str], |
| 3524 | degraded_by_source: dict[str, str] | None = None, |
| 3525 | ) -> list[str]: |
| 3526 | warnings: list[str] = [] |
| 3527 | if not candidates: |
| 3528 | warnings.append("No candidates survived retrieval and ranking.") |
| 3529 | if len(candidates) < 5: |
| 3530 | warnings.append("Evidence is thin for this topic.") |
| 3531 | top_sources = { |
| 3532 | source |
| 3533 | for candidate in candidates[:5] |
| 3534 | for source in schema.candidate_sources(candidate) |
| 3535 | } |
| 3536 | if len(top_sources) <= 1 and len(candidates) >= 3: |
| 3537 | warnings.append("Top evidence is highly concentrated in one source.") |
| 3538 | if errors_by_source: |
| 3539 | warnings.append(f"Some sources failed: {', '.join(sorted(errors_by_source))}") |
| 3540 | if degraded_by_source: |
| 3541 | # Partial failures: the source returned some items but errored/timed out |
| 3542 | # on at least one subquery, so its coverage is likely incomplete. Kept |
| 3543 | # distinct from hard failures so the signal is not silently dropped. |
| 3544 | warnings.append( |
| 3545 | f"Some sources returned partial results (degraded): {', '.join(sorted(degraded_by_source))}" |
| 3546 | ) |
| 3547 | if not items_by_source: |
| 3548 | warnings.append("No source returned usable items.") |
| 3549 | return warnings |
| 3550 | |
| 3551 | |
| 3552 | _STATUS_PREFIX = r"\b(?:https?(?:/\d(?:\.\d)?)?(?:\s+error)?|status(?:[\s_]*code)?|code)\s*[:=#]?\s*" |
| 3553 | |
| 3554 | |
| 3555 | def _mentions_status(msg: str, code_pattern: str, phrases: tuple[str, ...]) -> bool: |
| 3556 | """True when ``msg`` names an HTTP status matching ``code_pattern``. |
| 3557 | |
| 3558 | A bare number is not enough (``"batch 429 failed"`` is not a rate |
| 3559 | limit): the code must follow an HTTP/status/code marker, or co-occur |
| 3560 | with one of ``phrases`` (``"rate limited (429)"``). Digit-aware |
| 3561 | boundaries keep ``"14293"`` from reading as a 429. |
| 3562 | """ |
| 3563 | if not msg: |
| 3564 | return False |
| 3565 | code = r"(?:" + code_pattern + r")(?!\d)" |
| 3566 | if re.search(_STATUS_PREFIX + code, msg, re.IGNORECASE): |
| 3567 | return True |
| 3568 | lowered = msg.lower() |
| 3569 | return any(p in lowered for p in phrases) and re.search(r"(?<!\d)" + code, msg) is not None |
| 3570 | |
| 3571 | |
| 3572 | _RATE_LIMIT_PHRASES = ("rate limit", "rate-limit", "ratelimit", "too many requests") |
| 3573 | _SERVER_ERROR_PHRASES = ( |
| 3574 | "server error", |
| 3575 | "bad gateway", |
| 3576 | "service unavailable", |
| 3577 | "gateway timeout", |
| 3578 | "gateway time-out", |
| 3579 | ) |
| 3580 | |
| 3581 | |
| 3582 | def _is_rate_limit_error(exc: Exception) -> bool: |
| 3583 | """Detect 429 rate-limit errors by status code or message text.""" |
| 3584 | if hasattr(exc, "status_code") and getattr(exc, "status_code", None) == 429: |
| 3585 | return True |
| 3586 | return _mentions_status(str(exc), "429", _RATE_LIMIT_PHRASES) |
| 3587 | |
| 3588 | |
| 3589 | class SourceRunError(RuntimeError): |
| 3590 | """Source-specific failure that survived a module's fallback logic.""" |
| 3591 | |
| 3592 | def __init__(self, message: str, state: schema.RunOutcomeState | None = None): |
| 3593 | super().__init__(message) |
| 3594 | self.outcome_state = state or http.classify_failure(message=message) |
| 3595 | |
| 3596 | |
| 3597 | def _classify_source_failure(exc: Exception) -> tuple[schema.RunOutcomeState, bool]: |
| 3598 | """Classify HTTP, subprocess, and module-specific failures consistently.""" |
| 3599 | detail = str(exc) |
| 3600 | lowered = detail.lower() |
| 3601 | if any(marker in lowered for marker in ("not configured", "no api key", "not installed")): |
| 3602 | return schema.SKIPPED_UNCONFIGURED, False |
| 3603 | if any( |
| 3604 | marker in lowered |
| 3605 | for marker in ( |
| 3606 | "cookie expired", |
| 3607 | "expired cookie", |
| 3608 | "login required", |
| 3609 | "not logged in", |
| 3610 | "grok session expired", |
| 3611 | "session expired or was revoked", |
| 3612 | "invalid_grant", |
| 3613 | "not signed in", |
| 3614 | ) |
| 3615 | ): |
| 3616 | return schema.AUTH_FAILED, True |
| 3617 | state = getattr(exc, "outcome_state", None) or http.classify_failure( |
| 3618 | status_code=getattr(exc, "status_code", None), |
| 3619 | message=detail, |
| 3620 | ) |
| 3621 | return state, True |
| 3622 | |
| 3623 | |
| 3624 | def _outcome_artifact( |
| 3625 | state: schema.RunOutcomeState, |
| 3626 | detail: str, |
| 3627 | *, |
| 3628 | attempted: bool = True, |
| 3629 | ) -> dict[str, Any]: |
| 3630 | return { |
| 3631 | "_source_outcome": { |
| 3632 | "state": state, |
| 3633 | "detail": detail, |
| 3634 | "attempted": attempted, |
| 3635 | } |
| 3636 | } |
| 3637 | |
| 3638 | |
| 3639 | def _result_outcome_artifact(source: str, result: Any) -> dict[str, Any]: |
| 3640 | """Convert a legacy ``{"error": ...}`` source result into typed status.""" |
| 3641 | if not isinstance(result, dict) or not result.get("error"): |
| 3642 | return {} |
| 3643 | detail = str(result["error"]) |
| 3644 | if source == "reddit": |
| 3645 | state = reddit.classify_run_failure(detail) |
| 3646 | attempted = True |
| 3647 | elif source == "youtube": |
| 3648 | state = youtube_yt.classify_run_failure(detail) |
| 3649 | attempted = state != schema.SKIPPED_UNCONFIGURED |
| 3650 | elif source == "x": |
| 3651 | state = bird_x.classify_run_failure(detail) |
| 3652 | attempted = True |
| 3653 | elif source == "truthsocial" and detail == "Truth Social token expired": |
| 3654 | state = schema.AUTH_FAILED |
| 3655 | attempted = True |
| 3656 | elif source == "bluesky" and "network-level block" in detail.lower(): |
| 3657 | state = schema.UNREACHABLE |
| 3658 | attempted = True |
| 3659 | else: |
| 3660 | state, attempted = _classify_source_failure(SourceRunError(detail)) |
| 3661 | return _outcome_artifact(state, detail, attempted=attempted) |
| 3662 | |
| 3663 | |
| 3664 | def _legacy_artifact_outcome( |
| 3665 | source: str, |
| 3666 | artifact: Any, |
| 3667 | ) -> dict[str, Any] | None: |
| 3668 | """Map known pre-outcome artifact contracts to a typed outcome note.""" |
| 3669 | if not isinstance(artifact, dict): |
| 3670 | return None |
| 3671 | explicit = artifact.get("_source_outcome") |
| 3672 | if isinstance(explicit, dict): |
| 3673 | return explicit |
| 3674 | if source == "perplexity": |
| 3675 | candidates: list[tuple[str | None, dict[str, Any]]] = [(None, artifact)] |
| 3676 | if artifact.get("mode") == "both": |
| 3677 | for leg in ("search", "agent"): |
| 3678 | value = artifact.get(leg) |
| 3679 | if isinstance(value, dict): |
| 3680 | candidates.append((leg, value)) |
| 3681 | outcomes: list[dict[str, Any]] = [] |
| 3682 | for leg, candidate in candidates: |
| 3683 | if not candidate.get("error"): |
| 3684 | continue |
| 3685 | error = str(candidate["error"]) |
| 3686 | detail = str( |
| 3687 | candidate.get("backgroundErrorMessage") |
| 3688 | or candidate.get("backgroundPollError") |
| 3689 | or candidate.get("agentErrorMessage") |
| 3690 | or candidate.get("asyncErrorMessage") |
| 3691 | or candidate.get("message") |
| 3692 | or error |
| 3693 | ) |
| 3694 | if leg: |
| 3695 | detail = f"{leg} leg: {detail}" |
| 3696 | status_code = candidate.get("statusCode") |
| 3697 | if status_code is None: |
| 3698 | status_code = candidate.get("backgroundPollStatusCode") |
| 3699 | state = ( |
| 3700 | health.TIMEOUT |
| 3701 | if error.lower() == "timeout" |
| 3702 | else http.classify_failure( |
| 3703 | status_code=status_code, |
| 3704 | message=f"{error}: {detail}", |
| 3705 | ) |
| 3706 | ) |
| 3707 | outcomes.append(_outcome_artifact(state, detail)["_source_outcome"]) |
| 3708 | if outcomes: |
| 3709 | return min( |
| 3710 | outcomes, |
| 3711 | key=lambda outcome: _FAILURE_SPECIFICITY.get(outcome["state"], 9), |
| 3712 | ) |
| 3713 | if ( |
| 3714 | source == "grounding" |
| 3715 | and artifact.get("reason") == "keyless-search-unavailable" |
| 3716 | ): |
| 3717 | return _outcome_artifact( |
| 3718 | schema.UNREACHABLE, |
| 3719 | "Keyless web search unavailable", |
| 3720 | )["_source_outcome"] |
| 3721 | return None |
| 3722 | |
| 3723 | |
| 3724 | def _summarize_lane_failures(failures: list[http.HTTPError], source: str = "") -> str: |
| 3725 | """One line naming what a source lost to swallowed sub-request failures. |
| 3726 | |
| 3727 | ``"3 sub-requests rate-limited (HTTP 429); 1 sub-request blocked (HTTP 403)"``. |
| 3728 | Used as ``SourceOutcome.detail`` on a source that still delivered items, |
| 3729 | so the loss is visible to ``doctor --postmortem`` without branding the |
| 3730 | source partial (issue #985 wording; PR #959 semantics). |
| 3731 | """ |
| 3732 | counts: dict[tuple[str, int | None], int] = {} |
| 3733 | for failure in failures: |
| 3734 | state = getattr(failure, "outcome_state", None) or health.ERROR |
| 3735 | code = getattr(failure, "status_code", None) |
| 3736 | counts[(state, code)] = counts.get((state, code), 0) + 1 |
| 3737 | labels = { |
| 3738 | health.RATE_LIMITED: "rate-limited", |
| 3739 | health.AUTH_FAILED: "blocked", |
| 3740 | health.PAYMENT_REQUIRED: health.credits_exhausted_label(source), |
| 3741 | health.TIMEOUT: "timed out", |
| 3742 | health.UNREACHABLE: "unreachable", |
| 3743 | health.SCHEMA_DRIFT: "returned an unexpected shape", |
| 3744 | } |
| 3745 | parts = [] |
| 3746 | for (state, code), n in sorted(counts.items(), key=lambda kv: -kv[1]): |
| 3747 | noun = "sub-request" if n == 1 else "sub-requests" |
| 3748 | label = labels.get(state, "failed") |
| 3749 | suffix = f" (HTTP {code})" if code else "" |
| 3750 | parts.append(f"{n} {noun} {label}{suffix}") |
| 3751 | return "; ".join(parts) |
| 3752 | |
| 3753 | |
| 3754 | def _resolve_stream_outcome( |
| 3755 | source: str, |
| 3756 | artifact: Any, |
| 3757 | failures: list[http.HTTPError], |
| 3758 | ) -> dict[str, Any] | None: |
| 3759 | """Choose the most specific artifact or captured HTTP outcome.""" |
| 3760 | artifact_outcome = _legacy_artifact_outcome(source, artifact) |
| 3761 | if not failures: |
| 3762 | return artifact_outcome |
| 3763 | # Pick the most specific failure rather than the last-appended one: |
| 3764 | # parallel workers append in nondeterministic order, and an auth failure |
| 3765 | # must not be masked by a later 429 (wrong doctor prescription). |
| 3766 | failure = min( |
| 3767 | failures, |
| 3768 | key=lambda f: _FAILURE_SPECIFICITY.get(f.outcome_state, 9), |
| 3769 | ) |
| 3770 | captured_outcome = _outcome_artifact( |
| 3771 | failure.outcome_state, |
| 3772 | str(failure), |
| 3773 | )["_source_outcome"] |
| 3774 | if artifact_outcome is None: |
| 3775 | return captured_outcome |
| 3776 | if ( |
| 3777 | artifact_outcome.get("state") == health.ERROR |
| 3778 | and failure.outcome_state != health.ERROR |
| 3779 | ): |
| 3780 | return captured_outcome |
| 3781 | return artifact_outcome |
| 3782 | |
| 3783 | |
| 3784 | def _finalize_source_status( |
| 3785 | outcomes: dict[str, schema.SourceOutcome], |
| 3786 | items_by_source: dict[str, list[schema.SourceItem]], |
| 3787 | ) -> dict[str, schema.SourceOutcome]: |
| 3788 | """Sync outcome counts to the final post-filter evidence set.""" |
| 3789 | finalized: dict[str, schema.SourceOutcome] = {} |
| 3790 | for source, outcome in outcomes.items(): |
| 3791 | count = len(items_by_source.get(source, [])) |
| 3792 | state = outcome.state |
| 3793 | detail = outcome.detail |
| 3794 | fix_hint = outcome.fix_hint |
| 3795 | if state == schema.NO_RESULTS and count: |
| 3796 | state = health.OK |
| 3797 | detail = None |
| 3798 | fix_hint = None |
| 3799 | elif state == health.OK and not count: |
| 3800 | state = outcome.lane_failure_state or schema.NO_RESULTS |
| 3801 | elif state == schema.PARTIAL and not count: |
| 3802 | state = http.classify_failure(message=detail or "") |
| 3803 | finalized[source] = schema.SourceOutcome( |
| 3804 | source=source, |
| 3805 | state=state, |
| 3806 | items_returned=count, |
| 3807 | attempted=outcome.attempted, |
| 3808 | detail=detail, |
| 3809 | at=outcome.at, |
| 3810 | fix_hint=fix_hint, |
| 3811 | lane_failure_state=outcome.lane_failure_state, |
| 3812 | ) |
| 3813 | return finalized |
| 3814 | |
| 3815 | |
| 3816 | def _is_transient_error(exc: Exception) -> bool: |
| 3817 | """Detect 5xx server errors that are worth retrying.""" |
| 3818 | status = getattr(exc, "status_code", None) |
| 3819 | if isinstance(status, int) and 500 <= status < 600: |
| 3820 | return True |
| 3821 | return _mentions_status(str(exc), r"5\d\d", _SERVER_ERROR_PHRASES) |
| 3822 | |
| 3823 | |
| 3824 | def _topic_handle_mentions(topic: str) -> set[str]: |
| 3825 | """@mentions in the topic, which are real X handles. |
| 3826 | |
| 3827 | These are used to determine whether the subject was identified: an |
| 3828 | @mention like "@steipete" is a real handle that can exempt its owner from |
| 3829 | the relevance floor. Regular words like "Peter" are not real handles. |
| 3830 | """ |
| 3831 | return { |
| 3832 | mention.lower() |
| 3833 | for mention in re.findall(r"@([A-Za-z0-9_]{1,15})", topic or "") |
| 3834 | } |
| 3835 | |
| 3836 | |
| 3837 | def _topic_first_party_candidates(topic: str) -> set[str]: |
| 3838 | """Handle-shaped tokens in the topic itself, usable before any retrieval. |
| 3839 | |
| 3840 | Phase 1 runs before automatic handle resolution, and a quick-depth run |
| 3841 | skips that resolution entirely, so neither has access to the extracted |
| 3842 | handle set. Without this a quick search for "Peter Steinberger steipete" |
| 3843 | still drops every post steipete wrote, which is the exact failure this |
| 3844 | branch exists to fix. |
| 3845 | |
| 3846 | Deliberately permissive about what looks like a handle and strict about |
| 3847 | what it does: a candidate only ever matters if a retrieved post's *author* |
| 3848 | matches it, so an ordinary word like "lunch" costs nothing -- no author is |
| 3849 | named "lunch". The realistic false positive is an account named after a |
| 3850 | topic word, which the frequency-ranked path could surface anyway. |
| 3851 | """ |
| 3852 | tokens = set() |
| 3853 | for mention in re.findall(r"@([A-Za-z0-9_]{1,15})", topic or ""): |
| 3854 | tokens.add(mention.lower()) |
| 3855 | for word in re.findall(r"[A-Za-z0-9_]{3,15}", topic or ""): |
| 3856 | lowered = word.lower() |
| 3857 | if lowered not in relevance.STOPWORDS: |
| 3858 | tokens.add(lowered) |
| 3859 | return tokens |
| 3860 | |
| 3861 | |
| 3862 | def _name_lane_subject(topic: str) -> str: |
| 3863 | """Resolve the entity name to search for by name, not the whole topic. |
| 3864 | |
| 3865 | Phrase-quoting a raw topic ("Peter Steinberger steipete") matches nothing |
| 3866 | on X: nobody writes the handle and the display name together. Prefer a |
| 3867 | title-cased proper noun the way the planner's keyword query does, and fall |
| 3868 | back to the first compound term, then to the topic. |
| 3869 | """ |
| 3870 | import re as _re |
| 3871 | compounds = query.extract_compound_terms(topic) or [] |
| 3872 | title_cased = [ |
| 3873 | term for term in compounds |
| 3874 | if _re.match(r"^(?:[A-Z][a-z]+\s+){1,}[A-Z][a-z]+$", term) |
| 3875 | ] |
| 3876 | if title_cased: |
| 3877 | return title_cased[0] |
| 3878 | if compounds: |
| 3879 | return compounds[0] |
| 3880 | return topic.strip() |
| 3881 | |
| 3882 | |
| 3883 | def _run_supplemental_searches( |
| 3884 | *, |
| 3885 | topic: str, |
| 3886 | bundle: schema.RetrievalBundle, |
| 3887 | plan: schema.QueryPlan, |
| 3888 | config: dict[str, Any], |
| 3889 | depth: str, |
| 3890 | date_range: tuple[str, str], |
| 3891 | runtime: schema.ProviderRuntime, |
| 3892 | mock: bool, |
| 3893 | rate_limited_sources: set[str], |
| 3894 | rate_limit_lock: threading.Lock, |
| 3895 | x_handle: str | None = None, |
| 3896 | x_related: list[str] | None = None, |
| 3897 | resolved_handles_out: list[str] | None = None, |
| 3898 | ) -> None: |
| 3899 | """Phase 2: extract entities from Phase 1 results, run targeted supplemental searches.""" |
| 3900 | # The sanitized plan already intersects requested, available, and excluded |
| 3901 | # sources. A handle or host envelope must not expand that source boundary. |
| 3902 | if not any("x" in subquery.sources for subquery in plan.subqueries): |
| 3903 | return |
| 3904 | from_date, to_date = date_range |
| 3905 | |
| 3906 | # Host-fetched X lane: the envelope's lane calls replace the backend |
| 3907 | # lanes and are served at every depth (the host already paid for them), |
| 3908 | # before the quick/mock return and before the chain is recomputed. |
| 3909 | # Extracted-handle promotion is skipped on envelope runs; a declared lane |
| 3910 | # without an envelope runs no lane at all (the topic stream already |
| 3911 | # recorded the not-passed outcome). |
| 3912 | if config.get("_x_lane_missing"): |
| 3913 | return |
| 3914 | envelope = config.get("_x_envelope") |
| 3915 | if envelope is not None: |
| 3916 | _serve_envelope_lanes( |
| 3917 | envelope, bundle=bundle, plan=plan, x_handle=x_handle, x_related=x_related, |
| 3918 | from_date=from_date, to_date=to_date, |
| 3919 | ) |
| 3920 | return |
| 3921 | |
| 3922 | if depth == "quick" or mock: |
| 3923 | return |
| 3924 | |
| 3925 | # Convert SourceItems to dicts for entity_extract. All X items (whatever |
| 3926 | # backend fetched them — bird, xai, xurl, xquik) land under the single "x" |
| 3927 | # slug, so this reads the whole X corpus. |
| 3928 | x_dicts = [ |
| 3929 | {"author_handle": item.author or "", "text": item.body or ""} |
| 3930 | for item in bundle.items_by_source.get("x", []) |
| 3931 | ] |
| 3932 | reddit_dicts = [ |
| 3933 | { |
| 3934 | "subreddit": item.container or "", |
| 3935 | "comment_insights": item.metadata.get("comment_insights", []), |
| 3936 | "top_comments": [ |
| 3937 | {"excerpt": c.get("excerpt", c.get("text", ""))} |
| 3938 | for c in (item.metadata.get("top_comments") or []) |
| 3939 | if isinstance(c, dict) |
| 3940 | ], |
| 3941 | } |
| 3942 | for item in bundle.items_by_source.get("reddit", []) |
| 3943 | ] |
| 3944 | |
| 3945 | if not x_dicts and not reddit_dicts and not x_handle and not x_related: |
| 3946 | return |
| 3947 | |
| 3948 | entities = entity_extract.extract_entities( |
| 3949 | reddit_dicts, x_dicts, |
| 3950 | max_handles=3, max_subreddits=3, |
| 3951 | ) |
| 3952 | |
| 3953 | handles = entities.get("x_handles", []) |
| 3954 | |
| 3955 | # Add explicit --x-handle if provided |
| 3956 | if x_handle: |
| 3957 | handle_clean = x_handle.lstrip("@").lower() |
| 3958 | if handle_clean not in [h.lower() for h in handles]: |
| 3959 | handles.insert(0, handle_clean) |
| 3960 | |
| 3961 | # Collect related handles (searched separately with lower weight) |
| 3962 | related_handles = [] |
| 3963 | if x_related: |
| 3964 | primary_lower = x_handle.lstrip("@").lower() if x_handle else "" |
| 3965 | for rh in x_related: |
| 3966 | rh_clean = rh.lstrip("@").lower().strip() |
| 3967 | if rh_clean and rh_clean != primary_lower and rh_clean not in [h.lower() for h in handles]: |
| 3968 | related_handles.append(rh_clean) |
| 3969 | |
| 3970 | # Surface every handle this run resolved back to the caller. resolved_handles |
| 3971 | # is built later from --x-handle / --github-user / --x-related only, so |
| 3972 | # without this an auto-discovered subject handle never reaches it and every |
| 3973 | # downstream first-party protection (entity-miss exemption, FIRST_PARTY_FLOOR, |
| 3974 | # interaction floor) stays inert on any run that did not pass --x-handle. |
| 3975 | # Populated before the early return below so a run whose lanes cannot execute |
| 3976 | # still contributes its resolved handles. |
| 3977 | if resolved_handles_out is not None: |
| 3978 | # Only corroborated handles get first-party status. The extracted set is |
| 3979 | # frequency-ranked over retrieved post text, so a prolific commentator -- |
| 3980 | # or an engagement-farming account that posts on every topic -- lands in |
| 3981 | # it without being the subject. First-party status is strong: it exempts |
| 3982 | # an author from the relevance floor entirely and raises their per-author |
| 3983 | # cap, so granting it on frequency alone would let a spam account buy |
| 3984 | # immunity from filtering. Require the handle to look like the topic's |
| 3985 | # subject, or to have been named explicitly by the user. |
| 3986 | explicit = { |
| 3987 | h.lstrip("@").strip().lower() |
| 3988 | for h in ([x_handle] + list(x_related or [])) |
| 3989 | if h and h.strip() |
| 3990 | } |
| 3991 | topic_tokens = {t for t in re.findall(r"[a-z0-9]+", topic.lower()) if len(t) > 2} |
| 3992 | seen = {h.lower() for h in resolved_handles_out} |
| 3993 | for h in [*handles, *related_handles]: |
| 3994 | clean = h.lstrip("@").strip().lower() |
| 3995 | if not clean or clean in seen: |
| 3996 | continue |
| 3997 | corroborated = clean in explicit or any( |
| 3998 | token in clean or clean in token for token in topic_tokens |
| 3999 | ) |
| 4000 | if corroborated: |
| 4001 | resolved_handles_out.append(clean) |
| 4002 | seen.add(clean) |
| 4003 | |
| 4004 | if not handles and not related_handles: |
| 4005 | return |
| 4006 | |
| 4007 | # Pick the X handle-search backend: the first handle-capable backend in the |
| 4008 | # chain (grok, bird, xapi, or xquik). These supplemental from:/mentions lanes are |
| 4009 | # complementary to the topic search, so when the topic primary can't run |
| 4010 | # them (xai/xurl have no handle-lane implementation) but a capable backend |
| 4011 | # is available, use it rather than skipping Phase 2. bird scrapes X GraphQL |
| 4012 | # with the user's browser cookies; xquik runs the same lanes over its REST |
| 4013 | # API. All items land under the single "x" slug. |
| 4014 | x_slug = "x" |
| 4015 | |
| 4016 | def _record_handle_lane_failures(lane: str, failures: list[str]) -> bool: |
| 4017 | """Surface handle-lane failures the adapter would otherwise swallow. |
| 4018 | |
| 4019 | bird and xquik report a per-handle failure by returning no items, so an |
| 4020 | empty lane is indistinguishable from a subject who simply did not post. |
| 4021 | Left unrecorded, the run reports the X source as a clean zero and the |
| 4022 | report states as fact that nothing was posted. |
| 4023 | |
| 4024 | Returns True when the failure is auth-shaped, so the caller's existing |
| 4025 | AUTH_FAILED branch owns the message and the fix hint. Everything else |
| 4026 | (timeout, spawn failure, non-zero exit, bad JSON) is recorded here. |
| 4027 | ``record_failure`` keeps already-returned items as partial. |
| 4028 | """ |
| 4029 | if not failures: |
| 4030 | return False |
| 4031 | detail = "; ".join(failures[:3]) |
| 4032 | if len(failures) > 3: |
| 4033 | detail += f" (+{len(failures) - 3} more)" |
| 4034 | # Only xquik reports an auth-shaped handle-lane failure; bird's are all |
| 4035 | # transport (timeout / spawn / exit / JSON). Match the two phrases |
| 4036 | # xquik._execute_search actually emits rather than adding a fifth |
| 4037 | # copy of this repo's auth-marker vocabulary. |
| 4038 | lowered = detail.lower() |
| 4039 | if "auth failed" in lowered or "key unpaid" in lowered: |
| 4040 | return True |
| 4041 | bundle.record_failure( |
| 4042 | x_slug, |
| 4043 | health.UNREACHABLE, |
| 4044 | f"Phase 2 {lane}-lane: {detail}", |
| 4045 | attempted=True, |
| 4046 | ) |
| 4047 | return False |
| 4048 | |
| 4049 | chain = env.x_backend_chain(config) |
| 4050 | # Trust an explicit runtime backend as the head of the chain. |
| 4051 | pinned = runtime.x_search_backend |
| 4052 | if pinned: |
| 4053 | chain = [pinned] + [b for b in chain if b != pinned] |
| 4054 | primary = next((b for b in chain if b in ("grok", "bird", "xapi", "xquik")), None) |
| 4055 | |
| 4056 | # Name lane (posts naming the subject in plain text, no @-mention) is |
| 4057 | # grok-only for now: it needs phrase-quoting and negation operators the |
| 4058 | # other handle-capable backends do not expose uniformly. It is NOT a |
| 4059 | # fallback for the mention lane -- most discussion of a person or company |
| 4060 | # never @-mentions them, so the two lanes reach disjoint sets. |
| 4061 | _name_lane = None |
| 4062 | |
| 4063 | if primary == "grok": |
| 4064 | # One budget shared by all three lanes, started here rather than per |
| 4065 | # lane: the point is to bound the total, not each part. |
| 4066 | lane_deadline = time.monotonic() + grok_x.LANE_BUDGET_SECONDS |
| 4067 | |
| 4068 | def _from_lane(hs: list, count: int, and_topic: bool = False) -> tuple[list, bool]: |
| 4069 | items, revoked = grok_x.search_handles( |
| 4070 | hs, topic, from_date, to_date, count_per=count, |
| 4071 | deadline=lane_deadline, and_topic=and_topic, |
| 4072 | ) |
| 4073 | return items, revoked |
| 4074 | |
| 4075 | def _about_lane(hs: list, count: int) -> tuple[list, bool]: |
| 4076 | items, revoked = grok_x.search_mentions( |
| 4077 | hs, from_date, to_date, topic=topic, count_per=count, |
| 4078 | deadline=lane_deadline, |
| 4079 | ) |
| 4080 | return items, revoked |
| 4081 | |
| 4082 | def _name_lane(hs: list, count: int) -> tuple[list, bool]: |
| 4083 | # Use the resolved entity name, not the raw topic. Phrase-quoting |
| 4084 | # the whole topic ("Peter Steinberger steipete") matches nothing on |
| 4085 | # X; the subject's name is what other people actually write. |
| 4086 | subject = _name_lane_subject(topic) |
| 4087 | if not subject.strip(): |
| 4088 | return [], False |
| 4089 | items, revoked = grok_x.search_name( |
| 4090 | subject, from_date, to_date, exclude_handles=hs, count_per=count, |
| 4091 | deadline=lane_deadline, |
| 4092 | ) |
| 4093 | return items, revoked |
| 4094 | elif primary == "bird": |
| 4095 | def _from_lane(hs: list, count: int, and_topic: bool = False) -> tuple[list, bool]: |
| 4096 | # bird_x.search_handles doesn't support and_topic yet |
| 4097 | failures: list[str] = [] |
| 4098 | items = bird_x.search_handles( |
| 4099 | hs, topic, from_date, count_per=count, failure_out=failures, |
| 4100 | to_date=to_date, |
| 4101 | ) |
| 4102 | return items, _record_handle_lane_failures("FROM", failures) |
| 4103 | |
| 4104 | def _about_lane(hs: list, count: int) -> tuple[list, bool]: |
| 4105 | failures: list[str] = [] |
| 4106 | items = bird_x.search_mentions( |
| 4107 | hs, from_date, count_per=count, failure_out=failures, |
| 4108 | to_date=to_date, |
| 4109 | ) |
| 4110 | return items, _record_handle_lane_failures("ABOUT", failures) |
| 4111 | elif primary == "xapi": |
| 4112 | # Direct X API v2 with the app-only bearer: from:/@ lanes run over |
| 4113 | # search/all with the recent-search fallback. One budget shared by |
| 4114 | # every lane below (same shape as the grok lanes): a slow key bounds |
| 4115 | # the whole supplemental phase, not each call. |
| 4116 | xapi_token = config.get("X_BEARER_TOKEN") or "" |
| 4117 | xapi_deadline = time.monotonic() + x_api.LANE_BUDGET_SECONDS |
| 4118 | |
| 4119 | def _xapi_lane_receipt(lane_warnings: list[str]) -> None: |
| 4120 | # A deadline stop is incomplete coverage, reported in |
| 4121 | # report.warnings (the x_partial_coverage artifact), never a |
| 4122 | # healthy-looking silence and never a source failure. |
| 4123 | sink = bundle.artifacts.setdefault("x_partial_coverage", []) |
| 4124 | for note in lane_warnings: |
| 4125 | line = f"X handle lanes: {note}" |
| 4126 | if line not in sink: |
| 4127 | sink.append(line) |
| 4128 | |
| 4129 | def _from_lane(hs: list, count: int, and_topic: bool = False) -> tuple[list, bool]: |
| 4130 | # x_api.search_handles doesn't support and_topic; topic ranks only |
| 4131 | lane_warnings: list[str] = [] |
| 4132 | items = x_api.search_handles( |
| 4133 | hs, topic, from_date, to_date, count_per=count, token=xapi_token, |
| 4134 | deadline=xapi_deadline, warnings=lane_warnings, |
| 4135 | ) |
| 4136 | _xapi_lane_receipt(lane_warnings) |
| 4137 | return items, False |
| 4138 | |
| 4139 | def _about_lane(hs: list, count: int) -> tuple[list, bool]: |
| 4140 | lane_warnings: list[str] = [] |
| 4141 | items = x_api.search_mentions( |
| 4142 | hs, from_date, to_date, topic=topic, count_per=count, token=xapi_token, |
| 4143 | deadline=xapi_deadline, warnings=lane_warnings, |
| 4144 | ) |
| 4145 | _xapi_lane_receipt(lane_warnings) |
| 4146 | return items, False |
| 4147 | elif primary == "xquik": |
| 4148 | xquik_token = env.get_xquik_token(config) |
| 4149 | |
| 4150 | def _from_lane(hs: list, count: int, and_topic: bool = False) -> tuple[list, bool]: |
| 4151 | # xquik.search_handles doesn't support and_topic yet |
| 4152 | failures: list[str] = [] |
| 4153 | items = xquik.search_handles( |
| 4154 | hs, topic, from_date, to_date, count_per=count, token=xquik_token, |
| 4155 | failure_out=failures, |
| 4156 | ) |
| 4157 | return items, _record_handle_lane_failures("FROM", failures) |
| 4158 | |
| 4159 | def _about_lane(hs: list, count: int) -> tuple[list, bool]: |
| 4160 | failures: list[str] = [] |
| 4161 | items = xquik.search_mentions( |
| 4162 | hs, from_date, to_date, topic=topic, count_per=count, |
| 4163 | token=xquik_token, failure_out=failures, |
| 4164 | ) |
| 4165 | return items, _record_handle_lane_failures("ABOUT", failures) |
| 4166 | else: |
| 4167 | return # primary X backend has no handle-lane support (xai/xurl) or none configured |
| 4168 | |
| 4169 | # Skip if the X source is rate-limited. |
| 4170 | if x_slug in rate_limited_sources: |
| 4171 | return |
| 4172 | |
| 4173 | # Collect existing URLs for deduplication |
| 4174 | existing_urls = { |
| 4175 | item.url |
| 4176 | for items in bundle.items_by_source.values() |
| 4177 | for item in items |
| 4178 | if item.url |
| 4179 | } |
| 4180 | |
| 4181 | ranking_query = plan.subqueries[0].ranking_query if plan.subqueries else topic |
| 4182 | primary_label = plan.subqueries[0].label if plan.subqueries else "primary" |
| 4183 | |
| 4184 | # Split FROM promotion: determine which handles get FROM lane and how. |
| 4185 | # - Primary explicit handle (--x-handle): always FROM, no AND topic, full weight |
| 4186 | # - x_related handles: searched separately with lower weight (0.3), kept in |
| 4187 | # related_handles variable for the supplemental-related section below |
| 4188 | # - Extracted handles: FROM only if ≥2 on-topic hits AND ratio ≥0.5, |
| 4189 | # and those pulls DO AND the topic (from:handle Rome) |
| 4190 | primary_explicit = [x_handle] if x_handle else [] |
| 4191 | |
| 4192 | explicit_promotable, extracted_promotable = x_judge.promotable_handles( |
| 4193 | x_dicts, # Phase 1 X items for judging |
| 4194 | topic, |
| 4195 | handles, # entity_extract handles |
| 4196 | explicit_handles=primary_explicit, |
| 4197 | ranking_query=ranking_query, |
| 4198 | ) |
| 4199 | |
| 4200 | # All promotable handles for ABOUT and NAME lanes (primary only, not related) |
| 4201 | all_promotable = list(set(explicit_promotable + extracted_promotable)) |
| 4202 | |
| 4203 | # Search primary handles (full weight): FROM lane (their own tweets) + |
| 4204 | # ABOUT lane (tweets mentioning them). Both engagement-weighted and deduped |
| 4205 | # by URL at normalize time. |
| 4206 | any_revoked = False # Track auth revocation across lanes |
| 4207 | if all_promotable: |
| 4208 | # Independent try/except per lane so a failure in one does not discard |
| 4209 | # the other's already-computed results. |
| 4210 | from_items: list = [] |
| 4211 | about_items: list = [] |
| 4212 | about_revoked = False |
| 4213 | name_revoked = False |
| 4214 | |
| 4215 | # FROM lane: explicit handles without AND topic (person posts omit their own name) |
| 4216 | if explicit_promotable: |
| 4217 | try: |
| 4218 | explicit_items, explicit_revoked = _from_lane(explicit_promotable, FROM_LANE_COUNT_PER, and_topic=False) |
| 4219 | from_items.extend(explicit_items) |
| 4220 | if explicit_revoked: |
| 4221 | any_revoked = True |
| 4222 | bundle.record_failure( |
| 4223 | x_slug, schema.AUTH_FAILED, |
| 4224 | f"Phase 2 FROM-lane (explicit): {primary} authentication failed (session expired, revoked, or key unpaid)", |
| 4225 | attempted=True, |
| 4226 | ) |
| 4227 | except Exception as exc: |
| 4228 | print(f"[Pipeline] Phase 2 FROM-lane (explicit) failed: {exc}", file=sys.stderr) |
| 4229 | state, attempted = _classify_source_failure(exc) |
| 4230 | bundle.record_failure( |
| 4231 | x_slug, state, f"Phase 2 FROM-lane (explicit): {exc}", attempted=attempted, |
| 4232 | ) |
| 4233 | |
| 4234 | # FROM lane: extracted handles WITH AND topic (from:handle Rome) |
| 4235 | if extracted_promotable: |
| 4236 | try: |
| 4237 | extracted_items, extracted_revoked = _from_lane(extracted_promotable, FROM_LANE_COUNT_PER, and_topic=True) |
| 4238 | from_items.extend(extracted_items) |
| 4239 | if extracted_revoked: |
| 4240 | any_revoked = True |
| 4241 | bundle.record_failure( |
| 4242 | x_slug, schema.AUTH_FAILED, |
| 4243 | f"Phase 2 FROM-lane (extracted): {primary} authentication failed (session expired, revoked, or key unpaid)", |
| 4244 | attempted=True, |
| 4245 | ) |
| 4246 | except Exception as exc: |
| 4247 | print(f"[Pipeline] Phase 2 FROM-lane (extracted) failed: {exc}", file=sys.stderr) |
| 4248 | state, attempted = _classify_source_failure(exc) |
| 4249 | bundle.record_failure( |
| 4250 | x_slug, state, f"Phase 2 FROM-lane (extracted): {exc}", attempted=attempted, |
| 4251 | ) |
| 4252 | if not bundle.items_by_source.get(x_slug): |
| 4253 | bundle.errors_by_source[x_slug] = f"Phase 2 FROM-lane: {exc}" |
| 4254 | |
| 4255 | try: |
| 4256 | about_items, about_revoked = _about_lane(all_promotable, MENTION_LANE_COUNT_PER) |
| 4257 | if about_revoked: |
| 4258 | any_revoked = True |
| 4259 | bundle.record_failure( |
| 4260 | x_slug, schema.AUTH_FAILED, |
| 4261 | f"Phase 2 ABOUT-lane: {primary} authentication failed (session expired, revoked, or key unpaid)", |
| 4262 | attempted=True, |
| 4263 | ) |
| 4264 | except Exception as exc: |
| 4265 | print(f"[Pipeline] Phase 2 ABOUT-lane search failed: {exc}", file=sys.stderr) |
| 4266 | state, attempted = _classify_source_failure(exc) |
| 4267 | bundle.record_failure( |
| 4268 | x_slug, |
| 4269 | state, |
| 4270 | f"Phase 2 ABOUT-lane: {exc}", |
| 4271 | attempted=attempted, |
| 4272 | ) |
| 4273 | name_items: list = [] |
| 4274 | if _name_lane is not None: |
| 4275 | try: |
| 4276 | name_items, name_revoked = _name_lane(all_promotable, MENTION_LANE_COUNT_PER) |
| 4277 | if name_revoked: |
| 4278 | any_revoked = True |
| 4279 | bundle.record_failure( |
| 4280 | x_slug, schema.AUTH_FAILED, |
| 4281 | f"Phase 2 NAME-lane: {primary} authentication failed (session expired, revoked, or key unpaid)", |
| 4282 | attempted=True, |
| 4283 | ) |
| 4284 | except Exception as exc: |
| 4285 | print(f"[Pipeline] Phase 2 NAME-lane search failed: {exc}", file=sys.stderr) |
| 4286 | state, attempted = _classify_source_failure(exc) |
| 4287 | bundle.record_failure( |
| 4288 | x_slug, state, f"Phase 2 NAME-lane: {exc}", attempted=attempted, |
| 4289 | ) |
| 4290 | |
| 4291 | raw_items = from_items + about_items + name_items |
| 4292 | |
| 4293 | # Partial coverage is a reportable outcome, not a normal result: a |
| 4294 | # report carrying only one side of an entity topic is incomplete, and |
| 4295 | # without this it looks indistinguishable from genuinely thin |
| 4296 | # discussion. |
| 4297 | if _name_lane is not None: |
| 4298 | empty = [ |
| 4299 | label for label, items in |
| 4300 | (("by", from_items), ("mention", about_items), ("name", name_items)) |
| 4301 | if not items |
| 4302 | ] |
| 4303 | if empty and len(empty) < 3: |
| 4304 | # A warning, not a source outcome. record_failure would set the |
| 4305 | # X source to PARTIAL, which is outside _STRICT_EXIT_OK_STATES |
| 4306 | # and would make wrappers using LAST30DAYS_STRICT_EXIT exit 3 on |
| 4307 | # runs that returned perfectly good X coverage. An empty lane is |
| 4308 | # common and legitimate: the name lane carries an engagement |
| 4309 | # floor and the mention lane is empty for most non-famous |
| 4310 | # handles. |
| 4311 | bundle.artifacts.setdefault("x_partial_coverage", []).append( |
| 4312 | f"X partial coverage: {', '.join(empty)} lane(s) returned " |
| 4313 | "nothing; the report may show only one side of this entity." |
| 4314 | ) |
| 4315 | |
| 4316 | if raw_items: |
| 4317 | # First-party handles: only primary explicit handle, not promoted commentators |
| 4318 | # (first-party exempts from relevance floor; granting to commentators |
| 4319 | # would let junk become un-prunable) |
| 4320 | first_party_for_normalize = list(set( |
| 4321 | h.lower().lstrip("@") for h in primary_explicit if h |
| 4322 | )) |
| 4323 | normalized = _normalize_score_dedupe( |
| 4324 | x_slug, raw_items, from_date, to_date, |
| 4325 | freshness_mode=plan.freshness_mode, |
| 4326 | ranking_query=ranking_query, |
| 4327 | first_party_handles=first_party_for_normalize, |
| 4328 | ) |
| 4329 | # Deduplicate against Phase 1 URLs |
| 4330 | normalized = [item for item in normalized if item.url not in existing_urls] |
| 4331 | if normalized: |
| 4332 | bundle.add_items(primary_label, x_slug, normalized) |
| 4333 | # Update existing URLs for related-handle dedup |
| 4334 | for item in normalized: |
| 4335 | if item.url: |
| 4336 | existing_urls.add(item.url) |
| 4337 | |
| 4338 | # Search related handles with lower weight (0.3) |
| 4339 | # Related handles are explicit (--x-related), so FROM without AND topic. |
| 4340 | if related_handles: |
| 4341 | try: |
| 4342 | raw_items, rel_revoked = _from_lane(related_handles, RELATED_HANDLE_COUNT_PER, and_topic=False) |
| 4343 | if rel_revoked: |
| 4344 | any_revoked = True |
| 4345 | bundle.record_failure( |
| 4346 | x_slug, schema.AUTH_FAILED, |
| 4347 | f"Phase 2 related handle search: {primary} authentication failed (session expired, revoked, or key unpaid)", |
| 4348 | attempted=True, |
| 4349 | ) |
| 4350 | except Exception as exc: |
| 4351 | print(f"[Pipeline] Phase 2 related handle search failed: {exc}", file=sys.stderr) |
| 4352 | state, attempted = _classify_source_failure(exc) |
| 4353 | bundle.record_failure( |
| 4354 | x_slug, |
| 4355 | state, |
| 4356 | f"Phase 2 related handle search: {exc}", |
| 4357 | attempted=attempted, |
| 4358 | ) |
| 4359 | raw_items = [] |
| 4360 | |
| 4361 | if raw_items: |
| 4362 | normalized = _normalize_score_dedupe( |
| 4363 | x_slug, raw_items, from_date, to_date, |
| 4364 | freshness_mode=plan.freshness_mode, |
| 4365 | ranking_query=ranking_query, |
| 4366 | first_party_handles=related_handles, |
| 4367 | ) |
| 4368 | # Deduplicate against all existing URLs (Phase 1 + primary handles) |
| 4369 | normalized = [item for item in normalized if item.url not in existing_urls] |
| 4370 | if normalized: |
| 4371 | # Use a separate subquery label with lower weight so RRF |
| 4372 | # scores related-handle results below primary results. |
| 4373 | bundle.add_items("supplemental-related", x_slug, normalized) |
| 4374 | # Register the supplemental-related label in the plan for fusion |
| 4375 | if not any(sq.label == "supplemental-related" for sq in plan.subqueries): |
| 4376 | plan.subqueries.append( |
| 4377 | schema.SubQuery( |
| 4378 | label="supplemental-related", |
| 4379 | search_query=", ".join(related_handles), |
| 4380 | ranking_query=ranking_query, |
| 4381 | sources=[x_slug], |
| 4382 | weight=0.3, |
| 4383 | ) |
| 4384 | ) |
| 4385 | |
| 4386 | |
| 4387 | def _retry_thin_sources( |
| 4388 | *, |
| 4389 | topic: str, |
| 4390 | bundle: schema.RetrievalBundle, |
| 4391 | plan: schema.QueryPlan, |
| 4392 | config: dict[str, Any], |
| 4393 | depth: str, |
| 4394 | date_range: tuple[str, str], |
| 4395 | runtime: schema.ProviderRuntime, |
| 4396 | mock: bool, |
| 4397 | rate_limited_sources: set[str], |
| 4398 | rate_limit_lock: threading.Lock, |
| 4399 | settings: dict[str, Any], |
| 4400 | web_backend: str = "auto", |
| 4401 | skip_sources: set[str] | None = None, |
| 4402 | subreddits: list[str] | None = None, |
| 4403 | tiktok_hashtags: list[str] | None = None, |
| 4404 | tiktok_creators: list[str] | None = None, |
| 4405 | ig_creators: list[str] | None = None, |
| 4406 | first_party_handles: Iterable[str] | None = None, |
| 4407 | first_party_by_source: Mapping[str, Iterable[str]] | None = None, |
| 4408 | run_started: float | None = None, |
| 4409 | ) -> None: |
| 4410 | """Retry sources with thin results using simplified core subject query.""" |
| 4411 | if depth == "quick": |
| 4412 | return |
| 4413 | |
| 4414 | planned_sources: list[str] = [] |
| 4415 | for subquery in plan.subqueries: |
| 4416 | for source in subquery.sources: |
| 4417 | if source not in planned_sources: |
| 4418 | planned_sources.append(source) |
| 4419 | _skip = (skip_sources or set()) | THIN_RETRY_EXEMPT |
| 4420 | thin_sources = [ |
| 4421 | source |
| 4422 | for source in planned_sources |
| 4423 | if len(bundle.items_by_source.get(source, [])) < 3 |
| 4424 | and source not in bundle.errors_by_source |
| 4425 | and source not in _skip |
| 4426 | ] |
| 4427 | |
| 4428 | if not thin_sources: |
| 4429 | return |
| 4430 | |
| 4431 | core = query.extract_core_subject(topic, max_words=3) |
| 4432 | if not core: |
| 4433 | return |
| 4434 | # Note: we intentionally do NOT skip when core == topic. For short topics |
| 4435 | # like "Kanye West", the 3-word core IS the topic — but the planner may |
| 4436 | # have sent a different (worse) query to the source. Retrying with the |
| 4437 | # raw core subject is still valuable. |
| 4438 | |
| 4439 | from_date, to_date = date_range |
| 4440 | |
| 4441 | # Create a retry subquery with the simplified core subject |
| 4442 | retry_subquery = schema.SubQuery( |
| 4443 | label="retry", |
| 4444 | search_query=core, |
| 4445 | ranking_query=f"What recent evidence from the last 30 days matters for {core}?", |
| 4446 | sources=thin_sources, |
| 4447 | weight=0.3, |
| 4448 | ) |
| 4449 | |
| 4450 | def _retry_one_source( |
| 4451 | source: str, |
| 4452 | ) -> tuple[str, list[schema.SourceItem], dict[str, Any] | None]: |
| 4453 | raw_items, artifact = _retrieve_stream( |
| 4454 | topic=topic, |
| 4455 | subquery=retry_subquery, |
| 4456 | source=source, |
| 4457 | config=config, |
| 4458 | depth=depth, |
| 4459 | date_range=date_range, |
| 4460 | runtime=runtime, |
| 4461 | mock=mock, |
| 4462 | rate_limited_sources=rate_limited_sources, |
| 4463 | rate_limit_lock=rate_limit_lock, |
| 4464 | web_backend=web_backend, |
| 4465 | raw_topic=topic, |
| 4466 | subreddits=subreddits, |
| 4467 | tiktok_hashtags=tiktok_hashtags, |
| 4468 | tiktok_creators=tiktok_creators, |
| 4469 | ig_creators=ig_creators, |
| 4470 | run_started=run_started, |
| 4471 | # Skip Amazon review enrichment here to avoid duplicate Bright Data |
| 4472 | # pulls for ASINs already enriched in Phase 1. Finalize will enrich |
| 4473 | # any genuinely new products that weren't in Phase 1. |
| 4474 | skip_amazon_enrichment=True, |
| 4475 | ) |
| 4476 | outcome_note = artifact.get("_source_outcome") if isinstance(artifact, dict) else None |
| 4477 | detail_note = artifact.get("_source_outcome_detail") if isinstance(artifact, dict) else None |
| 4478 | detail_state = artifact.get("_source_outcome_detail_state") if isinstance(artifact, dict) else None |
| 4479 | normalized = _normalize_score_dedupe( |
| 4480 | source, |
| 4481 | raw_items, |
| 4482 | from_date, |
| 4483 | to_date, |
| 4484 | freshness_mode=plan.freshness_mode, |
| 4485 | ranking_query=retry_subquery.ranking_query, |
| 4486 | first_party_handles=first_party_handles, |
| 4487 | first_party_by_source=first_party_by_source, |
| 4488 | # Match Phase 1: X defers its relevance floor until the run has |
| 4489 | # resolved handles. Applying it here would discard a subject- |
| 4490 | # authored post that does not repeat the subject's name, and the |
| 4491 | # later resolved-handle floor cannot recover a post that never |
| 4492 | # entered the bundle. |
| 4493 | defer_relevance_prune=(source == "x"), |
| 4494 | ) |
| 4495 | if source == "jobs": |
| 4496 | return source, normalized, outcome_note, (detail_note, detail_state) |
| 4497 | normalized = _apply_reddit_stream_keepers( |
| 4498 | source, normalized, settings["per_stream_limit"], topic |
| 4499 | ) |
| 4500 | return source, normalized, outcome_note, (detail_note, detail_state) |
| 4501 | |
| 4502 | retryable = [s for s in thin_sources if s not in rate_limited_sources] |
| 4503 | |
| 4504 | from concurrent.futures import ThreadPoolExecutor, as_completed |
| 4505 | with ThreadPoolExecutor(max_workers=min(4, len(retryable) or 1)) as executor: |
| 4506 | futures = {executor.submit(_retry_one_source, s): s for s in retryable} |
| 4507 | for future in as_completed(futures): |
| 4508 | source = futures[future] |
| 4509 | try: |
| 4510 | source, normalized, outcome_note, (detail_note, detail_state) = future.result() |
| 4511 | if outcome_note: |
| 4512 | bundle.record_failure( |
| 4513 | source, |
| 4514 | outcome_note["state"], |
| 4515 | outcome_note["detail"], |
| 4516 | attempted=outcome_note.get("attempted", True), |
| 4517 | ) |
| 4518 | if detail_note: |
| 4519 | bundle.record_detail(source, detail_note, state=detail_state) |
| 4520 | existing_urls = {item.url for item in bundle.items_by_source.get(source, []) if item.url} |
| 4521 | new_items = [item for item in normalized if item.url not in existing_urls] |
| 4522 | |
| 4523 | if new_items: |
| 4524 | primary_label = plan.subqueries[0].label if plan.subqueries else "primary" |
| 4525 | bundle.add_items(primary_label, source, new_items) |
| 4526 | except Exception as exc: |
| 4527 | print(f"[Pipeline] Retry failed for {source}: {type(exc).__name__}: {exc}", file=sys.stderr) |
| 4528 | state, attempted = _classify_source_failure(exc) |
| 4529 | bundle.record_failure( |
| 4530 | source, |
| 4531 | state, |
| 4532 | f"Simplified-query retry failed: {exc}", |
| 4533 | attempted=attempted, |
| 4534 | ) |
| 4535 | |
| 4536 | |
| 4537 | def _fetch_x_backend(backend, query, from_date, to_date, depth, config, warnings=None): |
| 4538 | """Fetch X items from a single backend. Returns (items, error_str). |
| 4539 | |
| 4540 | ``warnings``, when given, collects backend receipts that are not |
| 4541 | failures (xapi's "window truncated to 7 days" after the recent-search |
| 4542 | fallback) so the X branch can surface them as run artifacts. |
| 4543 | |
| 4544 | Backends are tried in priority order by the caller (env.x_backend_chain); |
| 4545 | a non-empty error_str signals a hard failure (auth/payment/etc.) so the |
| 4546 | caller can fail over to the next backend or surface the error honestly. |
| 4547 | |
| 4548 | For grok, auth_revoked signals mid-run session revocation: the error |
| 4549 | string includes "grok session expired" so _classify_source_failure maps |
| 4550 | it to AUTH_FAILED with a proper fix hint, distinct from "never signed in". |
| 4551 | |
| 4552 | The ``query`` parameter is the compiled search query - typically |
| 4553 | ``raw_topic or topic`` (like Reddit/YouTube), NOT the planner's |
| 4554 | ``search_query`` which may contain operator strings like "Rome Italy". |
| 4555 | """ |
| 4556 | if backend == "bird": |
| 4557 | result = bird_x.search_x(query, from_date, to_date, depth=depth) |
| 4558 | items = bird_x.parse_bird_response(result, query=query) |
| 4559 | elif backend == "grok": |
| 4560 | result = grok_x.search_x(query, from_date, to_date, depth=depth) |
| 4561 | items = result.get("items", []) if isinstance(result, dict) else [] |
| 4562 | if isinstance(result, dict) and result.get("auth_revoked"): |
| 4563 | err = result.get("error") or "grok session expired or was revoked" |
| 4564 | return items, f"grok: {err}" |
| 4565 | elif backend == "xai": |
| 4566 | model = config.get("LAST30DAYS_X_MODEL") or config.get("XAI_MODEL_PIN") or providers.XAI_DEFAULT |
| 4567 | result = xai_x.search_x(config["XAI_API_KEY"], model, query, from_date, to_date, depth=depth) |
| 4568 | items = xai_x.parse_x_response(result) |
| 4569 | elif backend == "xurl": |
| 4570 | result = xurl_x.search_x(query, depth=depth) |
| 4571 | items = xurl_x.parse_x_response(result, topic=query) |
| 4572 | elif backend == "xquik": |
| 4573 | result = xquik.search_xquik(query, from_date, to_date, depth=depth, token=env.get_xquik_token(config)) |
| 4574 | items = xquik.parse_xquik_response(result) |
| 4575 | elif backend == "xapi": |
| 4576 | result = x_api.search_x(config.get("X_BEARER_TOKEN") or "", query, from_date, to_date, depth=depth) |
| 4577 | items = result.get("items", []) if isinstance(result, dict) else [] |
| 4578 | warning = result.get("warning") if isinstance(result, dict) else None |
| 4579 | if warning: |
| 4580 | print(f"[X] xapi: {warning}", file=sys.stderr) |
| 4581 | if warnings is not None: |
| 4582 | warnings.append(f"X: xapi {warning}") |
| 4583 | else: |
| 4584 | return [], f"unknown X backend: {backend}" |
| 4585 | err = result.get("error") if isinstance(result, dict) else "" |
| 4586 | return items, (err or "") |
| 4587 | |
| 4588 | |
| 4589 | def _reddit_post_key(item: dict) -> str: |
| 4590 | """Stable per-thread dedupe key (base36 post id from the url/permalink).""" |
| 4591 | url = item.get("url") or item.get("permalink") or "" |
| 4592 | m = re.search(r"/comments/([A-Za-z0-9]+)", url) |
| 4593 | return m.group(1) if m else url |
| 4594 | |
| 4595 | |
| 4596 | def _merge_reddit_items(free: list[dict], sc: list[dict]) -> list[dict]: |
| 4597 | """Merge free + ScrapeCreators Reddit items, free first, deduped by post id. |
| 4598 | |
| 4599 | Used when the thinness-floor trigger backfills a thin free run with SC, so a |
| 4600 | thread present in both is never double-listed. |
| 4601 | """ |
| 4602 | merged = list(free) |
| 4603 | seen = {_reddit_post_key(it) for it in free} |
| 4604 | for it in sc: |
| 4605 | key = _reddit_post_key(it) |
| 4606 | if key and key not in seen: |
| 4607 | seen.add(key) |
| 4608 | merged.append(it) |
| 4609 | return merged |
| 4610 | |
| 4611 | |
| 4612 | def _retrieve_stream(*args, **kwargs) -> tuple[list[dict], dict]: |
| 4613 | """Run one stream and retain HTTP failures swallowed by source adapters.""" |
| 4614 | # run_started is passed through but not used here; it goes to _retrieve_stream_impl |
| 4615 | source = str(kwargs.get("source") or "") |
| 4616 | fixture_request = { |
| 4617 | "source": source, |
| 4618 | "topic": kwargs.get("topic") or "", |
| 4619 | "search_query": getattr(kwargs.get("subquery"), "search_query", ""), |
| 4620 | "date_range": list(kwargs.get("date_range") or ()), |
| 4621 | "depth": kwargs.get("depth") or "", |
| 4622 | } |
| 4623 | module_backed = source in { |
| 4624 | "reddit", |
| 4625 | "x", |
| 4626 | "youtube", |
| 4627 | "stocktwits", |
| 4628 | "digg", |
| 4629 | "arxiv", |
| 4630 | "techmeme", |
| 4631 | "trustpilot", |
| 4632 | "github", |
| 4633 | } |
| 4634 | if module_backed: |
| 4635 | matched, replayed = http.fixture_source_replay(fixture_request) |
| 4636 | if matched: |
| 4637 | return replayed[0], replayed[1] |
| 4638 | try: |
| 4639 | with http.capture_failures() as failures, \ |
| 4640 | http.fixture_module_capture(module_backed): |
| 4641 | items, artifact = _retrieve_stream_impl(*args, **kwargs) |
| 4642 | except Exception as exc: |
| 4643 | recorded_exc = exc |
| 4644 | if failures and not getattr(exc, "outcome_state", None): |
| 4645 | failure = min( |
| 4646 | failures, |
| 4647 | key=lambda f: _FAILURE_SPECIFICITY.get(f.outcome_state, 9), |
| 4648 | ) |
| 4649 | recorded_exc = SourceRunError(str(exc), failure.outcome_state) |
| 4650 | if module_backed: |
| 4651 | http.fixture_source_record_error(fixture_request, recorded_exc) |
| 4652 | if recorded_exc is not exc: |
| 4653 | raise recorded_exc from exc |
| 4654 | raise |
| 4655 | outcome_note = _resolve_stream_outcome( |
| 4656 | str(kwargs.get("source") or ""), |
| 4657 | artifact, |
| 4658 | failures, |
| 4659 | ) |
| 4660 | if outcome_note: |
| 4661 | # Lane-level HTTP failures (e.g. a blocked shreddit partial on a |
| 4662 | # datacenter IP) are captured by the sink even when the source |
| 4663 | # delivered items. Only attach them when the run produced nothing, |
| 4664 | # or when the impl attached its own explicit outcome artifact (e.g. |
| 4665 | # "primary failed; fallback returned N items"). A swallowed lane |
| 4666 | # failure must not brand a successful source auth-failed/partial. |
| 4667 | # An adapter-declared outcome (typed ``_source_outcome`` or a legacy |
| 4668 | # ``{"error": ...}`` / per-leg artifact) is explicit and always |
| 4669 | # brands the source, even with items; only failures the adapter |
| 4670 | # swallowed into the capture sink are demoted to detail. |
| 4671 | explicit = isinstance(artifact, dict) and ( |
| 4672 | bool(artifact.get("_source_outcome")) |
| 4673 | or _legacy_artifact_outcome(str(kwargs.get("source") or ""), artifact) is not None |
| 4674 | ) |
| 4675 | if explicit or not items: |
| 4676 | artifact = dict(artifact or {}) |
| 4677 | artifact["_source_outcome"] = outcome_note |
| 4678 | elif failures: |
| 4679 | # The source delivered items. Keep it ``ok`` but carry what the |
| 4680 | # swallowed sub-requests lost, so doctor can still show it, and |
| 4681 | # the most specific failure state so a later empty filter result |
| 4682 | # or the thin-source retry can act on it. |
| 4683 | # Append to any detail the impl already set (e.g. a Reddit |
| 4684 | # backfill note) rather than overwriting it. |
| 4685 | artifact = dict(artifact or {}) |
| 4686 | summary = _summarize_lane_failures(failures, str(kwargs.get("source") or "")) |
| 4687 | prior = artifact.get("_source_outcome_detail") |
| 4688 | artifact["_source_outcome_detail"] = f"{prior}; {summary}" if prior else summary |
| 4689 | states = [f.outcome_state for f in failures] |
| 4690 | if artifact.get("_source_outcome_detail_state"): |
| 4691 | states.append(artifact["_source_outcome_detail_state"]) |
| 4692 | artifact["_source_outcome_detail_state"] = min( |
| 4693 | states, key=lambda state: _FAILURE_SPECIFICITY.get(state, 9) |
| 4694 | ) |
| 4695 | if module_backed: |
| 4696 | http.fixture_source_record(fixture_request, [items, artifact]) |
| 4697 | return items, artifact |
| 4698 | |
| 4699 | |
| 4700 | def _serve_envelope_topic(envelope: x_envelope.Envelope) -> tuple[list[dict], dict]: |
| 4701 | """Serve the envelope's topic-lane rows once. |
| 4702 | |
| 4703 | The first X subquery takes the rows and the envelope-status outcome; |
| 4704 | every later call (a second planner subquery, judge-retry, thin-retry) |
| 4705 | gets no items and no error, and no backend is ever consulted. |
| 4706 | """ |
| 4707 | items = envelope.take_topic() |
| 4708 | if items is None: |
| 4709 | return [], {} |
| 4710 | artifact: dict[str, Any] = {} |
| 4711 | if envelope.warnings: |
| 4712 | # A narrower host window is a receipt (report.warnings), not a failure. |
| 4713 | artifact["x_receipts"] = [f"X: {warning}" for warning in envelope.warnings] |
| 4714 | outcome = envelope.outcome() |
| 4715 | if outcome is not None: |
| 4716 | state, detail = outcome |
| 4717 | artifact.update(_outcome_artifact(state, detail)) |
| 4718 | return items, artifact |
| 4719 | |
| 4720 | |
| 4721 | def _serve_envelope_lanes( |
| 4722 | envelope: x_envelope.Envelope, |
| 4723 | *, |
| 4724 | bundle: schema.RetrievalBundle, |
| 4725 | plan: schema.QueryPlan, |
| 4726 | x_handle: str | None, |
| 4727 | x_related: list[str] | None, |
| 4728 | from_date: str, |
| 4729 | to_date: str, |
| 4730 | ) -> None: |
| 4731 | """Serve the envelope's from/mention/related calls into the lane merge. |
| 4732 | |
| 4733 | Mirrors the backend lanes: primary-handle rows (from + mention) join the |
| 4734 | primary subquery with first-party handling for the explicit handle and |
| 4735 | the per-handle lane counts; related rows join ``supplemental-related`` |
| 4736 | at the 0.3 weight. Lane claims were already validated at read time. |
| 4737 | """ |
| 4738 | calls = envelope.take_lanes() |
| 4739 | if not calls: |
| 4740 | return |
| 4741 | x_slug = "x" |
| 4742 | existing_urls = { |
| 4743 | item.url |
| 4744 | for items in bundle.items_by_source.values() |
| 4745 | for item in items |
| 4746 | if item.url |
| 4747 | } |
| 4748 | ranking_query = plan.subqueries[0].ranking_query if plan.subqueries else "" |
| 4749 | primary_label = plan.subqueries[0].label if plan.subqueries else "primary" |
| 4750 | primary_handles = sorted( |
| 4751 | {x_handle.lstrip("@").strip().lower()} if x_handle and x_handle.strip() else set() |
| 4752 | ) |
| 4753 | related_handles = [ |
| 4754 | h.lstrip("@").strip().lower() |
| 4755 | for h in (x_related or []) |
| 4756 | if h.strip() and h.lstrip("@").strip().lower() not in primary_handles |
| 4757 | ] |
| 4758 | |
| 4759 | def _cap_per_author(posts: list[dict], cap: int) -> list[dict]: |
| 4760 | seen: Counter[str] = Counter() |
| 4761 | kept: list[dict] = [] |
| 4762 | for post in posts: |
| 4763 | author = str(post.get("author_handle") or "").lower() |
| 4764 | if seen[author] >= cap: |
| 4765 | continue |
| 4766 | seen[author] += 1 |
| 4767 | kept.append(post) |
| 4768 | return kept |
| 4769 | |
| 4770 | primary_items: list[dict] = [] |
| 4771 | related_items: list[dict] = [] |
| 4772 | for call in calls: |
| 4773 | if call.lane == "from": |
| 4774 | primary_items.extend(_cap_per_author(call.posts, FROM_LANE_COUNT_PER)) |
| 4775 | elif call.lane == "mention": |
| 4776 | primary_items.extend( |
| 4777 | call.posts[: MENTION_LANE_COUNT_PER * max(1, len(call.handles))] |
| 4778 | ) |
| 4779 | elif call.lane == "related": |
| 4780 | related_items.extend(_cap_per_author(call.posts, RELATED_HANDLE_COUNT_PER)) |
| 4781 | |
| 4782 | if primary_items: |
| 4783 | normalized = _normalize_score_dedupe( |
| 4784 | x_slug, primary_items, from_date, to_date, |
| 4785 | freshness_mode=plan.freshness_mode, |
| 4786 | ranking_query=ranking_query, |
| 4787 | first_party_handles=primary_handles, |
| 4788 | ) |
| 4789 | normalized = [item for item in normalized if item.url not in existing_urls] |
| 4790 | if normalized: |
| 4791 | bundle.add_items(primary_label, x_slug, normalized) |
| 4792 | existing_urls.update(item.url for item in normalized if item.url) |
| 4793 | |
| 4794 | if related_items: |
| 4795 | normalized = _normalize_score_dedupe( |
| 4796 | x_slug, related_items, from_date, to_date, |
| 4797 | freshness_mode=plan.freshness_mode, |
| 4798 | ranking_query=ranking_query, |
| 4799 | first_party_handles=related_handles, |
| 4800 | ) |
| 4801 | normalized = [item for item in normalized if item.url not in existing_urls] |
| 4802 | if normalized: |
| 4803 | bundle.add_items("supplemental-related", x_slug, normalized) |
| 4804 | if not any(sq.label == "supplemental-related" for sq in plan.subqueries): |
| 4805 | plan.subqueries.append( |
| 4806 | schema.SubQuery( |
| 4807 | label="supplemental-related", |
| 4808 | search_query=", ".join(related_handles), |
| 4809 | ranking_query=ranking_query, |
| 4810 | sources=[x_slug], |
| 4811 | weight=0.3, |
| 4812 | ) |
| 4813 | ) |
| 4814 | |
| 4815 | |
| 4816 | def _retrieve_stream_impl( |
| 4817 | *, |
| 4818 | topic: str, |
| 4819 | subquery: schema.SubQuery, |
| 4820 | source: str, |
| 4821 | config: dict[str, Any], |
| 4822 | depth: str, |
| 4823 | date_range: tuple[str, str], |
| 4824 | runtime: schema.ProviderRuntime, |
| 4825 | mock: bool, |
| 4826 | rate_limited_sources: set[str] | None = None, |
| 4827 | rate_limit_lock: threading.Lock | None = None, |
| 4828 | web_backend: str = "auto", |
| 4829 | raw_topic: str = "", |
| 4830 | subreddits: list[str] | None = None, |
| 4831 | tiktok_hashtags: list[str] | None = None, |
| 4832 | tiktok_creators: list[str] | None = None, |
| 4833 | ig_creators: list[str] | None = None, |
| 4834 | trustpilot_domain: str | None = None, |
| 4835 | trustpilot_domain_is_hint: bool = False, |
| 4836 | run_started: float | None = None, |
| 4837 | skip_amazon_enrichment: bool = False, |
| 4838 | ) -> tuple[list[dict], dict]: |
| 4839 | # Early exit if source was rate-limited by a sibling future |
| 4840 | if rate_limited_sources is not None and source in rate_limited_sources: |
| 4841 | return [], {} |
| 4842 | from_date, to_date = date_range |
| 4843 | if mock: |
| 4844 | return _mock_stream_results(source, subquery) |
| 4845 | if source == "grounding": |
| 4846 | return grounding.web_search( |
| 4847 | subquery.search_query, date_range, config, backend=web_backend) |
| 4848 | if source == "jobs": |
| 4849 | return jobs.search_jobs( |
| 4850 | raw_topic or topic or subquery.search_query, |
| 4851 | date_range, |
| 4852 | config, |
| 4853 | depth=depth, |
| 4854 | web_backend=web_backend, |
| 4855 | explicit=bool(config.get("_hiring_signals_mode")), |
| 4856 | ) |
| 4857 | if source == "reddit": |
| 4858 | # Use raw_topic so expand_reddit_queries() generates diverse variants |
| 4859 | # from the original user topic, not the planner's narrowed search_query. |
| 4860 | reddit_query = raw_topic or subquery.search_query |
| 4861 | dedicated_subreddits = config.get("_dedicated_subreddits") or None |
| 4862 | has_sc_key = bool(config.get("SCRAPECREATORS_API_KEY")) |
| 4863 | sc_first = ( |
| 4864 | has_sc_key |
| 4865 | and (config.get(env.REDDIT_BACKEND_PIN_VAR) or "").lower() |
| 4866 | == "scrapecreators" |
| 4867 | ) |
| 4868 | if sc_first: |
| 4869 | # env.REDDIT_BACKEND_PIN_VAR=scrapecreators: SC primary, public fallback |
| 4870 | primary_failure: Exception | None = None |
| 4871 | try: |
| 4872 | result = reddit.search_and_enrich_memo( |
| 4873 | reddit_query, from_date, to_date, depth=depth, |
| 4874 | token=config.get("SCRAPECREATORS_API_KEY"), |
| 4875 | subreddits=subreddits, |
| 4876 | ) |
| 4877 | items = reddit.parse_reddit_response(result) |
| 4878 | if items: |
| 4879 | return items, {} |
| 4880 | sys.stderr.write( |
| 4881 | "[Reddit] ScrapeCreators primary returned no items, " |
| 4882 | "using public fallback\n" |
| 4883 | ) |
| 4884 | except Exception as exc: |
| 4885 | primary_failure = exc |
| 4886 | sys.stderr.write( |
| 4887 | f"[Reddit] ScrapeCreators primary failed " |
| 4888 | f"({type(exc).__name__}: {exc}), using public fallback\n" |
| 4889 | ) |
| 4890 | public_failure: Exception | None = None |
| 4891 | try: |
| 4892 | public_results = reddit_public.search_reddit_public( |
| 4893 | reddit_query, from_date, to_date, depth=depth, |
| 4894 | subreddits=subreddits, |
| 4895 | ) |
| 4896 | if public_results: |
| 4897 | if primary_failure is not None: |
| 4898 | state = reddit.classify_run_failure(str(primary_failure)) |
| 4899 | return public_results, _outcome_artifact( |
| 4900 | state, |
| 4901 | f"Reddit primary failed; public fallback returned " |
| 4902 | f"{len(public_results)} items: {primary_failure}", |
| 4903 | ) |
| 4904 | return public_results, {} |
| 4905 | sys.stderr.write( |
| 4906 | "[Reddit] Public fallback returned no items after " |
| 4907 | "ScrapeCreators primary miss\n" |
| 4908 | ) |
| 4909 | except Exception as exc: |
| 4910 | public_failure = exc |
| 4911 | sys.stderr.write( |
| 4912 | f"[Reddit] Public fallback also failed " |
| 4913 | f"({type(exc).__name__}: {exc})\n" |
| 4914 | ) |
| 4915 | failure = public_failure or primary_failure |
| 4916 | if failure is not None: |
| 4917 | state = reddit.classify_run_failure(str(failure)) |
| 4918 | raise SourceRunError( |
| 4919 | f"Reddit primary and fallback produced no results after failure: {failure}", |
| 4920 | state, |
| 4921 | ) |
| 4922 | return [], {} |
| 4923 | |
| 4924 | # Default: public Reddit first (free). ScrapeCreators backfills when the |
| 4925 | # free path returns fewer than the thinness floor (env.reddit_sc_min_items: |
| 4926 | # unset -> env.REDDIT_SC_MIN_ITEMS_DEFAULT, explicit 0 -> empty-only, |
| 4927 | # malformed -> 0 so a typo never spends extra credits). |
| 4928 | min_items = env.reddit_sc_min_items(config) |
| 4929 | public_results: list[dict] = [] |
| 4930 | public_failure: Exception | None = None |
| 4931 | try: |
| 4932 | public_results = reddit_public.search_reddit_public( |
| 4933 | reddit_query, from_date, to_date, depth=depth, |
| 4934 | subreddits=subreddits, dedicated_subreddits=dedicated_subreddits, |
| 4935 | ) or [] |
| 4936 | except Exception as exc: |
| 4937 | public_failure = exc |
| 4938 | sys.stderr.write( |
| 4939 | f"[Reddit] Public search failed ({type(exc).__name__}: {exc})" |
| 4940 | ) |
| 4941 | if not has_sc_key: |
| 4942 | sys.stderr.write("\n") |
| 4943 | state = reddit.classify_run_failure(str(exc)) |
| 4944 | raise SourceRunError(f"Reddit public search failed: {exc}", state) from exc |
| 4945 | sys.stderr.write(", using ScrapeCreators backup\n") |
| 4946 | # Enough free results, or no key to backfill with -> done. max(min_items, |
| 4947 | # 1) keeps the default (min_items=0) as empty-only AND treats exactly |
| 4948 | # `min_items` results as acceptable (no backfill) for min_items > 0. |
| 4949 | if len(public_results) >= max(min_items, 1) or not has_sc_key: |
| 4950 | return public_results, {} |
| 4951 | if public_results: |
| 4952 | sys.stderr.write( |
| 4953 | f"[Reddit] Free path returned {len(public_results)} " |
| 4954 | f"(below the {min_items}-item floor); backfilling with ScrapeCreators\n" |
| 4955 | ) |
| 4956 | # With free items in hand, scope the backfill's own HTTP failures |
| 4957 | # away from the stream's capture sink: a ScrapeCreators 429 is not a |
| 4958 | # reddit.com rate limit, and a RATE_LIMITED lane state would put |
| 4959 | # 'reddit' in rate_limited_sources and skip later Reddit streams and |
| 4960 | # the thin-source retry. Its failures become the detail note instead. |
| 4961 | # With no free items the backfill IS the source, so failures reach |
| 4962 | # the stream sink as before. |
| 4963 | backfill_scope = ( |
| 4964 | http.capture_failures() if public_results else contextlib.nullcontext([]) |
| 4965 | ) |
| 4966 | try: |
| 4967 | with backfill_scope as backfill_failures: |
| 4968 | result = reddit.search_and_enrich_memo( |
| 4969 | reddit_query, from_date, to_date, depth=depth, |
| 4970 | token=config.get("SCRAPECREATORS_API_KEY"), |
| 4971 | subreddits=subreddits, |
| 4972 | ) |
| 4973 | sc_items = reddit.parse_reddit_response(result) |
| 4974 | except Exception as exc: |
| 4975 | sys.stderr.write( |
| 4976 | f"[Reddit] ScrapeCreators backup also failed " |
| 4977 | f"({type(exc).__name__}: {exc})\n" |
| 4978 | ) |
| 4979 | state = reddit.classify_run_failure(str(exc)) |
| 4980 | if public_results: |
| 4981 | # The free path delivered: Reddit is working. The failed |
| 4982 | # backfill is a detail note, never the source's outcome. |
| 4983 | artifact = { |
| 4984 | "_source_outcome_detail": ( |
| 4985 | f"ScrapeCreators backfill failed after " |
| 4986 | f"{len(public_results)} free items: {exc}" |
| 4987 | ), |
| 4988 | } |
| 4989 | if state != health.RATE_LIMITED: |
| 4990 | # A ScrapeCreators rate limit says nothing about |
| 4991 | # reddit.com (and the run memo already stops a retry). |
| 4992 | artifact["_source_outcome_detail_state"] = state |
| 4993 | return public_results, artifact |
| 4994 | return public_results, _outcome_artifact( |
| 4995 | state, |
| 4996 | f"Reddit backup failed after {len(public_results)} public items: {exc}", |
| 4997 | ) |
| 4998 | merged = _merge_reddit_items(public_results, sc_items) |
| 4999 | if public_failure is not None: |
| 5000 | state = reddit.classify_run_failure(str(public_failure)) |
| 5001 | return merged, _outcome_artifact( |
| 5002 | state, |
| 5003 | f"Reddit public search failed; backup returned {len(sc_items)} items: " |
| 5004 | f"{public_failure}", |
| 5005 | ) |
| 5006 | trigger = ( |
| 5007 | f"below the {min_items}-item floor" if min_items > 0 else "free path empty" |
| 5008 | ) |
| 5009 | backfill_detail = ( |
| 5010 | f"ScrapeCreators backfill ran ({len(public_results)} free items, " |
| 5011 | f"{trigger}); added {len(merged) - len(public_results)} items" |
| 5012 | ) |
| 5013 | if backfill_failures: |
| 5014 | summary = _summarize_lane_failures(backfill_failures, "reddit") |
| 5015 | backfill_detail = f"{backfill_detail}; ScrapeCreators backfill lost {summary}" |
| 5016 | return merged, {"_source_outcome_detail": backfill_detail} |
| 5017 | if source == "x": |
| 5018 | if config.get("_x_lane_missing"): |
| 5019 | # The model declared the connector lane but passed no envelope. |
| 5020 | return [], _outcome_artifact(health.ERROR, x_envelope.DETAIL_NOT_PASSED) |
| 5021 | envelope = config.get("_x_envelope") |
| 5022 | if envelope is not None: |
| 5023 | # Host-fetched lane: the envelope replaces the backend chain and |
| 5024 | # is single-serve, so no backend runs and no judge-retry follows. |
| 5025 | return _serve_envelope_topic(envelope) |
| 5026 | |
| 5027 | # Compile X query from raw_topic (like Reddit/YouTube), not planner's |
| 5028 | # search_query which may contain operator strings like "Rome Italy". |
| 5029 | x_query = raw_topic or topic or subquery.search_query |
| 5030 | ranking_query = subquery.ranking_query |
| 5031 | |
| 5032 | # One X source, an ordered chain of interchangeable backends. Try the |
| 5033 | # primary; fall through to the next only if it returns nothing or errors. |
| 5034 | chain = env.x_backend_chain(config) |
| 5035 | # Trust an explicit runtime backend as the primary (already resolved as |
| 5036 | # available), keeping the rest of the chain as failover backups. |
| 5037 | pinned = runtime.x_search_backend |
| 5038 | if pinned: |
| 5039 | chain = [pinned] + [b for b in chain if b != pinned] |
| 5040 | if not chain: |
| 5041 | raise SourceRunError( |
| 5042 | "No X backend is available (not configured).", |
| 5043 | schema.SKIPPED_UNCONFIGURED, |
| 5044 | ) |
| 5045 | last_error = "" |
| 5046 | chain_errors: list[str] = [] |
| 5047 | items = [] |
| 5048 | used_backend = None |
| 5049 | x_warnings: list[str] = [] |
| 5050 | for i, backend in enumerate(chain): |
| 5051 | items, err = _fetch_x_backend( |
| 5052 | backend, x_query, from_date, to_date, depth, config, warnings=x_warnings, |
| 5053 | ) |
| 5054 | if items: |
| 5055 | if i > 0: |
| 5056 | # xapi is metered: name the spend when it served as a backup. |
| 5057 | spend = " (spends X API credits)" if backend == "xapi" else "" |
| 5058 | print( |
| 5059 | f"[X] primary backend(s) returned nothing; used fallback '{backend}'{spend}", |
| 5060 | file=sys.stderr, |
| 5061 | ) |
| 5062 | # Check for auth errors before proceeding to judge-retry |
| 5063 | if last_error: |
| 5064 | # Fallback succeeded after earlier backend failed. Classify |
| 5065 | # the original error: if it was AUTH_FAILED (grok revoked), |
| 5066 | # preserve that state so user gets re-login guidance. |
| 5067 | prior_state = http.classify_failure(message=last_error) |
| 5068 | if prior_state == schema.AUTH_FAILED: |
| 5069 | # Keep AUTH_FAILED visible so host shows re-login hint |
| 5070 | return items, _outcome_artifact( |
| 5071 | schema.AUTH_FAILED, |
| 5072 | f"X served via {backend} after {last_error}; re-login needed for primary backend", |
| 5073 | ) |
| 5074 | # Prior error was non-auth. Check if *current* backend also |
| 5075 | # reported an error (e.g., grok returned items + revocation). |
| 5076 | if err: |
| 5077 | current_state = http.classify_failure(message=err) |
| 5078 | if current_state == schema.AUTH_FAILED: |
| 5079 | return items, _outcome_artifact( |
| 5080 | schema.AUTH_FAILED, |
| 5081 | f"X served {len(items)} items via {backend} but also errored: {err}; re-login needed", |
| 5082 | ) |
| 5083 | # Non-auth prior error, no current auth error → fallback OK |
| 5084 | return items, _outcome_artifact( |
| 5085 | health.OK, |
| 5086 | f"X served via {backend} after {last_error}", |
| 5087 | ) |
| 5088 | if err: |
| 5089 | # Mixed result: backend returned items BUT also hit an error |
| 5090 | # (e.g., grok got some posts then auth was revoked mid-fanout). |
| 5091 | # Surface the error so the user gets re-login guidance. |
| 5092 | state = http.classify_failure(message=err) |
| 5093 | return items, _outcome_artifact( |
| 5094 | state, |
| 5095 | f"X returned {len(items)} items but also errored: {err}", |
| 5096 | ) |
| 5097 | # No auth issues and no prior errors - proceed to judge-retry |
| 5098 | used_backend = backend |
| 5099 | break |
| 5100 | if err: |
| 5101 | last_error = f"{backend}: {err}" |
| 5102 | chain_errors.append(last_error) |
| 5103 | print(f"[X] backend '{backend}' failed ({err}); trying next", file=sys.stderr) |
| 5104 | |
| 5105 | if not items and last_error: |
| 5106 | # A credit-exhaustion failure earlier in the chain is the most |
| 5107 | # specific outcome (top up, not re-authenticate); a later |
| 5108 | # backend's generic failure must not mask it. |
| 5109 | for candidate in chain_errors: |
| 5110 | if http.classify_failure(message=candidate) == health.PAYMENT_REQUIRED: |
| 5111 | last_error = candidate |
| 5112 | break |
| 5113 | state = ( |
| 5114 | bird_x.classify_run_failure(last_error) |
| 5115 | if last_error.startswith("bird:") |
| 5116 | else http.classify_failure(message=last_error) |
| 5117 | ) |
| 5118 | raise SourceRunError(f"All X backends failed — {last_error}", state) |
| 5119 | |
| 5120 | # Retrieve-judge-retry: judge corpus and retry if off-topic flood. |
| 5121 | # Skip retry on quick/mock (same as Phase 2). |
| 5122 | artifact = {} |
| 5123 | if x_warnings: |
| 5124 | # e.g. xapi's "window truncated to 7 days": a receipt that reaches |
| 5125 | # report.warnings (see the grounding artifacts walk in |
| 5126 | # _build_report), never a source failure. |
| 5127 | artifact["x_receipts"] = list(x_warnings) |
| 5128 | if items and depth != "quick" and not mock: |
| 5129 | items_for_judge = [ |
| 5130 | {"author_handle": it.get("author_handle", ""), "text": it.get("text", "")} |
| 5131 | for it in items |
| 5132 | ] |
| 5133 | if x_judge.should_retry_x_search(items_for_judge, x_query, ranking_query=ranking_query, depth=depth): |
| 5134 | # Retry with cleaned query (1 retry, ≤2 extra grok calls) |
| 5135 | # Strip noise words but preserve all significant terms to avoid |
| 5136 | # losing disambiguating terms (e.g., "react server components") |
| 5137 | core_tokens = query.extract_core_subject(x_query) |
| 5138 | retry_query = core_tokens or x_query |
| 5139 | print(f"[X] corpus off-topic; retrying with '{retry_query}'", file=sys.stderr) |
| 5140 | |
| 5141 | if used_backend: |
| 5142 | retry_items, retry_err = _fetch_x_backend( |
| 5143 | used_backend, retry_query, from_date, to_date, depth, config |
| 5144 | ) |
| 5145 | if retry_items: |
| 5146 | # Judge retry corpus |
| 5147 | retry_for_judge = [ |
| 5148 | {"author_handle": it.get("author_handle", ""), "text": it.get("text", "")} |
| 5149 | for it in retry_items |
| 5150 | ] |
| 5151 | retry_judgment = x_judge.judge_x_corpus( |
| 5152 | retry_for_judge, x_query, ranking_query=ranking_query |
| 5153 | ) |
| 5154 | orig_judgment = x_judge.judge_x_corpus( |
| 5155 | items_for_judge, x_query, ranking_query=ranking_query |
| 5156 | ) |
| 5157 | # Use retry if better on-topic ratio |
| 5158 | if retry_judgment["on_topic_ratio"] > orig_judgment["on_topic_ratio"]: |
| 5159 | print( |
| 5160 | f"[X] retry improved on-topic ratio: " |
| 5161 | f"{orig_judgment['on_topic_ratio']:.0%} -> " |
| 5162 | f"{retry_judgment['on_topic_ratio']:.0%}", |
| 5163 | file=sys.stderr, |
| 5164 | ) |
| 5165 | items = retry_items |
| 5166 | |
| 5167 | # Prune off-topic items before the pool. Eight on-topic → ok with 8. |
| 5168 | # Zero on-topic after retry → no-results, not ok with 40 junk. |
| 5169 | # Only prune items that have text to judge; items without text pass through. |
| 5170 | original_count = len(items) |
| 5171 | items_with_text = [(i, it) for i, it in enumerate(items) if it.get("text", "").strip()] |
| 5172 | |
| 5173 | if items_with_text: |
| 5174 | items_for_prune = [ |
| 5175 | {"author_handle": it.get("author_handle", ""), "text": it.get("text", "")} |
| 5176 | for _, it in items_with_text |
| 5177 | ] |
| 5178 | judgment = x_judge.judge_x_corpus( |
| 5179 | items_for_prune, x_query, ranking_query=ranking_query |
| 5180 | ) |
| 5181 | # Build set of indices for on-topic items |
| 5182 | on_topic_indices = set() |
| 5183 | for (orig_idx, _), pruned_item in zip(items_with_text, items_for_prune): |
| 5184 | if pruned_item in judgment["on_topic_items"]: |
| 5185 | on_topic_indices.add(orig_idx) |
| 5186 | |
| 5187 | # Keep items that are on-topic OR have no text (can't judge) |
| 5188 | items = [ |
| 5189 | it for i, it in enumerate(items) |
| 5190 | if i in on_topic_indices or not it.get("text", "").strip() |
| 5191 | ] |
| 5192 | |
| 5193 | # Record warning if significant pruning occurred (artifact, not failure) |
| 5194 | if len(items) < original_count: |
| 5195 | pruned = original_count - len(items) |
| 5196 | artifact.setdefault("_warnings", []).append( |
| 5197 | f"X: pruned {pruned} off-topic items; {len(items)} on-topic remain" |
| 5198 | ) |
| 5199 | |
| 5200 | if last_error and items: |
| 5201 | state = ( |
| 5202 | bird_x.classify_run_failure(last_error) |
| 5203 | if last_error.startswith("bird:") |
| 5204 | else http.classify_failure(message=last_error) |
| 5205 | ) |
| 5206 | return items, _outcome_artifact( |
| 5207 | state, |
| 5208 | f"X fallback '{used_backend}' returned {len(items)} items after {last_error}", |
| 5209 | ) |
| 5210 | return items, artifact |
| 5211 | if source == "youtube": |
| 5212 | # Use raw_topic so expand_youtube_queries() generates diverse variants |
| 5213 | # from the original user topic, not the planner's narrowed search_query. |
| 5214 | yt_query = raw_topic or subquery.search_query |
| 5215 | result = None |
| 5216 | youtube_failure: str | None = None |
| 5217 | # ScrapeCreators key (when present) is the default-on backup tier: it |
| 5218 | # powers the per-video transcript fallback, the SC search fallback, and |
| 5219 | # comment enrichment. None when no key, which keeps everything keyless. |
| 5220 | sc_token = ( |
| 5221 | config.get("SCRAPECREATORS_API_KEY", "") |
| 5222 | if env.is_youtube_sc_available(config) else None |
| 5223 | ) |
| 5224 | # Try yt-dlp first; the SC transcript fallback covers per-video failures. |
| 5225 | if which("yt-dlp"): |
| 5226 | try: |
| 5227 | result = youtube_yt.search_and_transcribe( |
| 5228 | yt_query, from_date, to_date, depth=depth, token=sc_token, |
| 5229 | ) |
| 5230 | if result.get("error"): |
| 5231 | youtube_failure = str(result["error"]) |
| 5232 | except Exception as exc: |
| 5233 | youtube_failure = str(exc) |
| 5234 | result = None |
| 5235 | # Fall back to SC YouTube search if yt-dlp failed or isn't installed. |
| 5236 | if (result is None or not result.get("items")) and sc_token: |
| 5237 | try: |
| 5238 | result = youtube_yt.search_youtube_sc( |
| 5239 | yt_query, from_date, to_date, depth=depth, token=sc_token, |
| 5240 | ) |
| 5241 | if result.get("error"): |
| 5242 | youtube_failure = str(result["error"]) |
| 5243 | except Exception as exc: |
| 5244 | youtube_failure = str(exc) |
| 5245 | result = None |
| 5246 | if result is None: |
| 5247 | result = {"items": []} |
| 5248 | # Enrich top videos with comments (default-on when a key is present). |
| 5249 | items = youtube_yt.parse_youtube_response(result) |
| 5250 | if items and env.is_youtube_comments_available(config): |
| 5251 | youtube_yt.enrich_with_comments( |
| 5252 | items, token=config.get("SCRAPECREATORS_API_KEY", ""), |
| 5253 | ) |
| 5254 | if youtube_failure: |
| 5255 | state = youtube_yt.classify_run_failure(youtube_failure) |
| 5256 | attempted = state != schema.SKIPPED_UNCONFIGURED |
| 5257 | return items, _outcome_artifact(state, youtube_failure, attempted=attempted) |
| 5258 | return items, {} |
| 5259 | if source == "tiktok": |
| 5260 | # Use raw_topic so expand_tiktok_queries() generates diverse variants |
| 5261 | # from the original user topic, not the planner's narrowed search_query. |
| 5262 | tiktok_query = raw_topic or subquery.search_query |
| 5263 | result = tiktok.search_and_enrich( |
| 5264 | tiktok_query, |
| 5265 | from_date, |
| 5266 | to_date, |
| 5267 | depth=depth, |
| 5268 | token=env.get_tiktok_token(config), |
| 5269 | hashtags=tiktok_hashtags, |
| 5270 | creators=tiktok_creators, |
| 5271 | ) |
| 5272 | items = tiktok.parse_tiktok_response(result) |
| 5273 | if items and env.is_tiktok_comments_available(config): |
| 5274 | sc_token = config.get("SCRAPECREATORS_API_KEY", "") |
| 5275 | tiktok.enrich_with_comments(items, token=sc_token) |
| 5276 | return items, _result_outcome_artifact(source, result) |
| 5277 | if source == "instagram": |
| 5278 | # Use raw_topic so expand_instagram_queries() generates diverse variants |
| 5279 | # from the original user topic, not the planner's narrowed search_query. |
| 5280 | ig_query = raw_topic or subquery.search_query |
| 5281 | result = instagram.search_and_enrich( |
| 5282 | ig_query, |
| 5283 | from_date, |
| 5284 | to_date, |
| 5285 | depth=depth, |
| 5286 | token=env.get_instagram_token(config), |
| 5287 | ig_creators=ig_creators, |
| 5288 | ) |
| 5289 | items = instagram.parse_instagram_response(result) |
| 5290 | if items and env.is_instagram_comments_available(config): |
| 5291 | instagram.enrich_with_comments( |
| 5292 | items, token=config.get("SCRAPECREATORS_API_KEY", ""), |
| 5293 | ) |
| 5294 | return items, _result_outcome_artifact(source, result) |
| 5295 | if source == "linkedin": |
| 5296 | token = config.get("SCRAPECREATORS_API_KEY", "") |
| 5297 | result = linkedin.search_linkedin( |
| 5298 | subquery.search_query, |
| 5299 | from_date, |
| 5300 | to_date, |
| 5301 | depth=depth, |
| 5302 | token=token, |
| 5303 | ) |
| 5304 | items = linkedin.parse_linkedin_response( |
| 5305 | result, from_date=from_date, to_date=to_date |
| 5306 | ) |
| 5307 | # Articles never appear in post search — surface them (high signal) |
| 5308 | # via a bounded profile-enrichment lane on person topics. |
| 5309 | items += linkedin.enrich_articles( |
| 5310 | items, raw_topic or topic, token, from_date=from_date, to_date=to_date |
| 5311 | ) |
| 5312 | return items, _result_outcome_artifact(source, result) |
| 5313 | if source == "hackernews": |
| 5314 | result = hackernews.search_hackernews(subquery.search_query, from_date, to_date, depth=depth) |
| 5315 | return ( |
| 5316 | hackernews.parse_hackernews_response(result, query=subquery.search_query), |
| 5317 | _result_outcome_artifact(source, result), |
| 5318 | ) |
| 5319 | if source == "stocktwits": |
| 5320 | # Pass raw_topic so symbol detection sees the full topic, not the |
| 5321 | # narrowed per-subquery search_query (same rationale as reddit). |
| 5322 | result = stocktwits.search_stocktwits( |
| 5323 | raw_topic or topic or subquery.search_query, from_date, to_date, depth=depth) |
| 5324 | return ( |
| 5325 | stocktwits.parse_stocktwits_response(result, query=subquery.search_query), |
| 5326 | _result_outcome_artifact(source, result), |
| 5327 | ) |
| 5328 | if source == "dripstack": |
| 5329 | result = dripstack.search_dripstack( |
| 5330 | subquery.search_query, from_date, to_date, depth=depth) |
| 5331 | relevance_topic = raw_topic or topic or subquery.search_query |
| 5332 | return ( |
| 5333 | dripstack.parse_dripstack_response(result, query=relevance_topic), |
| 5334 | _result_outcome_artifact(source, result), |
| 5335 | ) |
| 5336 | if source == "digg": |
| 5337 | result = digg.search_digg(subquery.search_query, from_date, to_date, depth=depth) |
| 5338 | items = digg.parse_digg_response(result, query=subquery.search_query) |
| 5339 | # Enrichment with attached X posts is deferred to |
| 5340 | # _finalize_items_by_source so it runs on the items that actually |
| 5341 | # survive dedupe rather than on top-K of the raw fanout. |
| 5342 | return items, _result_outcome_artifact(source, result) |
| 5343 | if source == "arxiv": |
| 5344 | result = arxiv.search_arxiv(subquery.search_query, from_date, to_date, depth=depth) |
| 5345 | # Relevance keys off the stable research topic, not the per-subquery |
| 5346 | # search_query, so off-topic narrowing does not let weak matches through. |
| 5347 | relevance_topic = raw_topic or topic or subquery.search_query |
| 5348 | return ( |
| 5349 | arxiv.parse_arxiv_response(result, query=relevance_topic), |
| 5350 | _result_outcome_artifact(source, result), |
| 5351 | ) |
| 5352 | if source == "techmeme": |
| 5353 | result = techmeme.search_techmeme(subquery.search_query, from_date, to_date, depth=depth) |
| 5354 | relevance_topic = raw_topic or topic or subquery.search_query |
| 5355 | return ( |
| 5356 | techmeme.parse_techmeme_response(result, query=relevance_topic), |
| 5357 | _result_outcome_artifact(source, result), |
| 5358 | ) |
| 5359 | if source == "trustpilot": |
| 5360 | # Brand-shape gate keys off the stable research topic, not the narrowed |
| 5361 | # per-subquery search_query, so the company is detected consistently. |
| 5362 | relevance_topic = raw_topic or topic or subquery.search_query |
| 5363 | result = trustpilot.search_trustpilot( |
| 5364 | relevance_topic, from_date, to_date, depth=depth, config=config, |
| 5365 | explicit_domain=trustpilot_domain, |
| 5366 | domain_is_hint=trustpilot_domain_is_hint, |
| 5367 | ) |
| 5368 | return ( |
| 5369 | trustpilot.parse_trustpilot_response(result, query=relevance_topic), |
| 5370 | _result_outcome_artifact(source, result), |
| 5371 | ) |
| 5372 | if source == "amazon": |
| 5373 | # The search keyword is model-supplied and may differ from the topic |
| 5374 | # ("Matt Van Horn" searches "June Oven"), so it keys off the stable |
| 5375 | # research topic rather than the narrowed per-subquery search_query. |
| 5376 | keyword = ( |
| 5377 | str((config or {}).get("_amazon_query") or "").strip() |
| 5378 | or raw_topic or topic or subquery.search_query |
| 5379 | ) |
| 5380 | domain = str((config or {}).get("LAST30DAYS_AMAZON_DOMAIN") or amazon.DEFAULT_DOMAIN) |
| 5381 | result = amazon.search_products(keyword, domain=domain, config=config) |
| 5382 | products = amazon.parse_search_response(result, keyword, domain=domain) |
| 5383 | artifact = _result_outcome_artifact(source, result) |
| 5384 | |
| 5385 | # Skip enrichment when called from thin retry (_retry_thin_sources) to |
| 5386 | # avoid duplicate Bright Data pulls for ASINs already enriched in Phase 1. |
| 5387 | # Finalize will enrich any NEW products (enrich_source_items no-ops when |
| 5388 | # top_comments is already set, so duplicates get skipped there too). |
| 5389 | if skip_amazon_enrichment: |
| 5390 | return products, artifact |
| 5391 | |
| 5392 | # Start review enrichment now, while other sources are still running. |
| 5393 | # Elapsed is measured from run_started so multi-source runs that finish |
| 5394 | # search quickly (30-90s) still have 190-250s of budget (clamped to 180). |
| 5395 | # This replaces the old deferred-to-finalize path which left only crumbs |
| 5396 | # (e.g. 11s) after long retrieval phases. |
| 5397 | elapsed = time.monotonic() - run_started if run_started else 0.0 |
| 5398 | enriched, review_status = amazon.enrich_with_reviews( |
| 5399 | products, |
| 5400 | depth=depth, |
| 5401 | config=config, |
| 5402 | elapsed=elapsed, |
| 5403 | keyword=keyword, |
| 5404 | ) |
| 5405 | |
| 5406 | # Record PARTIAL status if review lane was skipped or all pulls dropped |
| 5407 | if review_status: |
| 5408 | artifact = artifact or {} |
| 5409 | artifact = dict(artifact) if artifact else {} |
| 5410 | artifact["_source_outcome"] = { |
| 5411 | "state": schema.PARTIAL, |
| 5412 | "detail": review_status, |
| 5413 | "attempted": True, |
| 5414 | } |
| 5415 | |
| 5416 | return enriched, artifact |
| 5417 | if source == "meta_ads": |
| 5418 | # The advertiser is resolved from the stable research topic, not the |
| 5419 | # narrowed per-subquery search_query: a subquery like "kettle reviews" |
| 5420 | # would resolve a different page than the brand the run is about. |
| 5421 | brand = raw_topic or topic or subquery.search_query |
| 5422 | result = meta_ads.search_meta_ads( |
| 5423 | brand, |
| 5424 | from_date, |
| 5425 | to_date, |
| 5426 | depth=depth, |
| 5427 | token=(config or {}).get("SCRAPECREATORS_API_KEY") or "", |
| 5428 | country=str( |
| 5429 | (config or {}).get("LAST30DAYS_META_ADS_COUNTRY") |
| 5430 | or meta_ads.DEFAULT_COUNTRY |
| 5431 | ), |
| 5432 | page_override=str((config or {}).get("_meta_ads_page") or "").strip(), |
| 5433 | ) |
| 5434 | if result.get("partial"): |
| 5435 | # A partial lane carries `error` too, so the generic classifier |
| 5436 | # would run and have its verdict overwritten here regardless. |
| 5437 | artifact = { |
| 5438 | "_source_outcome": { |
| 5439 | "state": schema.PARTIAL, |
| 5440 | "detail": str(result.get("error") or "partial"), |
| 5441 | "attempted": True, |
| 5442 | } |
| 5443 | } |
| 5444 | else: |
| 5445 | artifact = dict(_result_outcome_artifact(source, result) or {}) |
| 5446 | # The footer needs the resolved advertiser and the pre-truncation |
| 5447 | # counts even on a run that produced zero items, and stream artifacts |
| 5448 | # only reach the report through the grounding list, so they ride here |
| 5449 | # and are lifted to top-level artifacts after retrieval. |
| 5450 | artifact["meta_ads_page"] = result.get("page") or {} |
| 5451 | artifact["meta_ads_tally"] = result.get("tally") or {} |
| 5452 | return result.get("ads") or [], artifact |
| 5453 | if source == "bluesky": |
| 5454 | result = bluesky.search_bluesky(subquery.search_query, from_date, to_date, depth=depth, config=config) |
| 5455 | return bluesky.parse_bluesky_response(result), _result_outcome_artifact(source, result) |
| 5456 | if source == "threads": |
| 5457 | result = threads.search_threads( |
| 5458 | subquery.search_query, from_date, to_date, |
| 5459 | depth=depth, |
| 5460 | token=config.get("SCRAPECREATORS_API_KEY"), |
| 5461 | ) |
| 5462 | return threads.parse_threads_response(result), _result_outcome_artifact(source, result) |
| 5463 | if source == "telegram": |
| 5464 | result = telegram.search_telegram( |
| 5465 | subquery.search_query, from_date, to_date, |
| 5466 | depth=depth, |
| 5467 | token=config.get("SCRAPECREATORS_API_KEY"), |
| 5468 | config=config, |
| 5469 | ) |
| 5470 | return telegram.parse_telegram_response(result), _result_outcome_artifact(source, result) |
| 5471 | if source == "truthsocial": |
| 5472 | result = truthsocial.search_truthsocial(subquery.search_query, from_date, to_date, depth=depth, config=config) |
| 5473 | return truthsocial.parse_truthsocial_response(result), _result_outcome_artifact(source, result) |
| 5474 | if source == "polymarket": |
| 5475 | result = polymarket.search_polymarket(subquery.search_query, from_date, to_date, depth=depth) |
| 5476 | # Relevance filtering keys off the stable original research topic, not the |
| 5477 | # per-subquery search_query (which narrows differently on each fanout pass |
| 5478 | # and would let off-topic markets through on broad subqueries while dropping |
| 5479 | # everything on narrow ones). |
| 5480 | relevance_topic = raw_topic or topic or subquery.search_query |
| 5481 | return ( |
| 5482 | polymarket.parse_polymarket_response(result, topic=relevance_topic), |
| 5483 | _result_outcome_artifact(source, result), |
| 5484 | ) |
| 5485 | if source == "github": |
| 5486 | # Resolve once at the pipeline boundary so search and enrich |
| 5487 | # share the result; otherwise each call would re-run the env |
| 5488 | # lookup and gh-CLI subprocess fallback (up to 5s timeout each). |
| 5489 | token = github.resolve_token(config.get("GITHUB_TOKEN")) |
| 5490 | response = github.search_github(subquery.search_query, from_date, to_date, depth=depth, token=token) |
| 5491 | items = github.parse_github_response(response) |
| 5492 | # Note: an unauth rate-limit (response["error"]) is expected on the |
| 5493 | # tokenless anon tier and returns empty here rather than raising — github |
| 5494 | # is now always eligible, so raising would spam "github failed" on every |
| 5495 | # tokenless run. The condition is logged in github.search_github. |
| 5496 | items = github.enrich_with_comments(items, depth=depth, token=token) |
| 5497 | return items, _result_outcome_artifact(source, response) |
| 5498 | if source == "pinterest": |
| 5499 | result = pinterest.search_pinterest( |
| 5500 | subquery.search_query, from_date, to_date, |
| 5501 | depth=depth, |
| 5502 | token=env.get_pinterest_token(config), |
| 5503 | ) |
| 5504 | return pinterest.parse_pinterest_response(result), _result_outcome_artifact(source, result) |
| 5505 | if source == "xiaohongshu": |
| 5506 | return xiaohongshu_api.search_feeds( |
| 5507 | subquery.search_query, |
| 5508 | from_date, |
| 5509 | to_date, |
| 5510 | env.get_xiaohongshu_api_base(config), |
| 5511 | depth=depth, |
| 5512 | ), {} |
| 5513 | if source == "perplexity": |
| 5514 | return perplexity.search(subquery.search_query, date_range, config, deep=config.get("_deep_research", False)) |
| 5515 | raise RuntimeError(f"Unsupported source: {source}") |
| 5516 | |
| 5517 | |
| 5518 | def _google_key(config: dict[str, Any]) -> str | None: |
| 5519 | return config.get("GOOGLE_API_KEY") or config.get("GEMINI_API_KEY") or config.get("GOOGLE_GENAI_API_KEY") |
| 5520 | |
| 5521 | |
| 5522 | |
| 5523 | |
| 5524 | def _mock_stream_results(source: str, subquery: schema.SubQuery) -> tuple[list[dict], dict]: |
| 5525 | # Namespace URLs and the canned comment by topic: real runs never hand two |
| 5526 | # distinct stories byte-identical evidence, and discovery's same-story fold |
| 5527 | # (correctly) collapses topics that share it. Mock enrichment sub-runs feed |
| 5528 | # this fixture one topic per subquery, so the slug keeps them distinct. |
| 5529 | slug = re.sub(r"[^a-z0-9]+", "-", subquery.search_query.lower()).strip("-") or "topic" |
| 5530 | payloads = { |
| 5531 | "reddit": [ |
| 5532 | { |
| 5533 | "id": "R1", |
| 5534 | "title": f"{subquery.search_query} discussion thread", |
| 5535 | "url": f"https://reddit.com/r/example/comments/{slug}-1", |
| 5536 | "subreddit": "example", |
| 5537 | "date": dates.get_date_range(5)[0], |
| 5538 | "engagement": {"score": 120, "num_comments": 48, "upvote_ratio": 0.91}, |
| 5539 | "selftext": f"Community discussion about {subquery.search_query}.", |
| 5540 | "top_comments": [{"excerpt": f"Strong firsthand feedback from {subquery.search_query} users."}], |
| 5541 | "relevance": 0.82, |
| 5542 | "why_relevant": "Mock Reddit result", |
| 5543 | } |
| 5544 | ], |
| 5545 | "x": [ |
| 5546 | { |
| 5547 | "id": "X1", |
| 5548 | "text": f"People on X are discussing {subquery.search_query} right now.", |
| 5549 | "url": f"https://x.com/example/status/{slug}-1", |
| 5550 | "author_handle": "example", |
| 5551 | "date": dates.get_date_range(2)[0], |
| 5552 | "engagement": {"likes": 200, "reposts": 35, "replies": 18, "quotes": 4}, |
| 5553 | "relevance": 0.79, |
| 5554 | "why_relevant": "Mock X result", |
| 5555 | } |
| 5556 | ], |
| 5557 | "grounding": [ |
| 5558 | { |
| 5559 | "id": "WB1", |
| 5560 | "title": f"{subquery.search_query} article", |
| 5561 | "url": f"https://example.com/article/{slug}", |
| 5562 | "source_domain": "example.com", |
| 5563 | "snippet": f"Recent web reporting about {subquery.search_query}.", |
| 5564 | "date": dates.get_date_range(7)[0], |
| 5565 | "relevance": 0.88, |
| 5566 | "why_relevant": "Brave web search", |
| 5567 | } |
| 5568 | ], |
| 5569 | "digg": [ |
| 5570 | { |
| 5571 | "id": "mock1abc", |
| 5572 | "title": f"Digg cluster about {subquery.search_query}", |
| 5573 | "url": f"https://di.gg/ai/mock1abc-{slug}", |
| 5574 | "tldr": f"Curated cluster summarizing recent {subquery.search_query} discussion across the AI 1000.", |
| 5575 | "author": "", |
| 5576 | "date": dates.get_date_range(3)[0], |
| 5577 | "engagement": {"postCount": 8, "uniqueAuthors": 5, "rank": 2, "rank_score": 49.0}, |
| 5578 | "first_post_age": "3d", |
| 5579 | "posts": [ |
| 5580 | { |
| 5581 | "username": "exampledev", |
| 5582 | "display_name": "Example Dev", |
| 5583 | "category": "Engineer", |
| 5584 | "rank": 142, |
| 5585 | "body": f"Quote from the AI 1000 about {subquery.search_query}.", |
| 5586 | "post_type": "tweet", |
| 5587 | "x_url": "https://x.com/exampledev/status/1", |
| 5588 | "posted_at": dates.get_date_range(3)[0], |
| 5589 | }, |
| 5590 | ], |
| 5591 | "relevance": 0.84, |
| 5592 | "why_relevant": "Mock Digg cluster", |
| 5593 | }, |
| 5594 | { |
| 5595 | "id": "mock2def", |
| 5596 | "title": f"Second Digg cluster on {subquery.search_query}", |
| 5597 | "url": f"https://di.gg/ai/mock2def-{slug}", |
| 5598 | "tldr": f"Another angle on {subquery.search_query}.", |
| 5599 | "author": "", |
| 5600 | "date": dates.get_date_range(8)[0], |
| 5601 | "engagement": {"postCount": 3, "uniqueAuthors": 2, "rank": 18, "rank_score": 33.0}, |
| 5602 | "first_post_age": "8d", |
| 5603 | "posts": [], |
| 5604 | "relevance": 0.71, |
| 5605 | "why_relevant": "Mock Digg cluster", |
| 5606 | }, |
| 5607 | ], |
| 5608 | "arxiv": [ |
| 5609 | { |
| 5610 | "id": f"http://arxiv.org/abs/2606.00001v1-{slug}", |
| 5611 | "title": f"A Survey of {subquery.search_query}", |
| 5612 | "url": f"https://arxiv.org/abs/2606.00001v1-{slug}", |
| 5613 | "summary": f"We present a comprehensive study of {subquery.search_query} and its recent advances.", |
| 5614 | "author": "Ada Lovelace et al.", |
| 5615 | "authors": ["Ada Lovelace", "Alan Turing"], |
| 5616 | "date": dates.get_date_range(20)[0], |
| 5617 | "engagement": {}, |
| 5618 | "relevance": 0.86, |
| 5619 | "why_relevant": "Mock arXiv paper", |
| 5620 | }, |
| 5621 | ], |
| 5622 | "techmeme": [ |
| 5623 | { |
| 5624 | "id": f"https://www.techmeme.com/260627/p1-{slug}", |
| 5625 | "title": f"Major development in {subquery.search_query} reshapes the industry", |
| 5626 | "url": f"https://www.techmeme.com/260627/p1-{slug}", |
| 5627 | "source_name": "techcrunch.com", |
| 5628 | "date": dates.get_date_range(1)[0], |
| 5629 | "engagement": {}, |
| 5630 | "relevance": 0.83, |
| 5631 | "why_relevant": "Mock Techmeme headline", |
| 5632 | }, |
| 5633 | ], |
| 5634 | "dripstack": [ |
| 5635 | { |
| 5636 | "id": "DS1", |
| 5637 | "title": f"Deep dive: {subquery.search_query} from a paid newsletter", |
| 5638 | "url": f"https://newsletter.example.com/deep-dive-{slug}", |
| 5639 | "author": "newsletter.example.com", |
| 5640 | "date": dates.get_date_range(3)[0], |
| 5641 | "engagement": {}, |
| 5642 | "relevance": 0.85, |
| 5643 | "why_relevant": "Mock DripStack newsletter result", |
| 5644 | "snippet": f"Professional analyst coverage of {subquery.search_query}.", |
| 5645 | "metadata": { |
| 5646 | "publication_slug": "newsletter.example.com", |
| 5647 | "post_slug": "deep-dive", |
| 5648 | "relevance_score": 85, |
| 5649 | "match_confidence": "strong", |
| 5650 | }, |
| 5651 | }, |
| 5652 | ], |
| 5653 | "trustpilot": [ |
| 5654 | { |
| 5655 | "id": "example.com", |
| 5656 | "title": f"{subquery.search_query}: TrustScore 3.4", |
| 5657 | "url": f"https://www.trustpilot.com/review/{slug}.example.com", |
| 5658 | "summary": f"Across recent reviews, customers were split on {subquery.search_query}: some praised support, others cited delays.", |
| 5659 | "name": subquery.search_query, |
| 5660 | "trustScore": 3.4, |
| 5661 | "reviewCount": 128, |
| 5662 | "date": dates.get_date_range(1)[0], |
| 5663 | "engagement": {"reviews": 128, "trustScore": 3.4}, |
| 5664 | "relevance": 0.8, |
| 5665 | "why_relevant": "Mock Trustpilot sentiment", |
| 5666 | }, |
| 5667 | ], |
| 5668 | # Three products spanning the drift states the footer renders: one |
| 5669 | # sagging below its all-time average (with enough in-window reviews |
| 5670 | # to clear the arrow threshold), one steady, and one too new to have |
| 5671 | # a baseline. Mock runs exercise the full R1c line without a CLI. |
| 5672 | "amazon": [ |
| 5673 | { |
| 5674 | "asin": "B0MOCK00X1", |
| 5675 | "date": dates.get_date_range(1)[1], |
| 5676 | "name": f"{subquery.search_query} Pro Model | Flagship Edition", |
| 5677 | "short_name": "Pro Model", |
| 5678 | "brand": subquery.search_query.split()[0].title() if subquery.search_query else "Example", |
| 5679 | "url": "https://www.amazon.com/dp/B0MOCK00X1", |
| 5680 | "rating": 4.4, |
| 5681 | "num_ratings": 459, |
| 5682 | "price": 39.99, |
| 5683 | "currency": "USD", |
| 5684 | "badge": "Best Seller", |
| 5685 | "sponsored": False, |
| 5686 | "relevance": 0.85, |
| 5687 | "why_relevant": "Mock Amazon product", |
| 5688 | "product_rating": 4.4, |
| 5689 | "product_rating_count": 459, |
| 5690 | "star_distribution": { |
| 5691 | "one_star": 28, "two_star": 9, "three_star": 28, |
| 5692 | "four_star": 60, "five_star": 335, |
| 5693 | }, |
| 5694 | "top_comments": [ |
| 5695 | { |
| 5696 | "score": 3, "rating": 2, "verified": True, |
| 5697 | "date": dates.get_date_range(3)[1], |
| 5698 | "excerpt": "The tray shifts in transit and the lid jams shut.", |
| 5699 | "title": "Lid jams", |
| 5700 | }, |
| 5701 | { |
| 5702 | "score": 1, "rating": 4, "verified": True, |
| 5703 | "date": dates.get_date_range(9)[1], |
| 5704 | "excerpt": "Solid build, but arrived with a dented panel.", |
| 5705 | "title": "Shipping dent", |
| 5706 | }, |
| 5707 | { |
| 5708 | "score": 0, "rating": 5, "verified": True, |
| 5709 | "date": dates.get_date_range(14)[1], |
| 5710 | "excerpt": "Keeps everything cold through a full school day.", |
| 5711 | "title": "Works great", |
| 5712 | }, |
| 5713 | { |
| 5714 | "score": 0, "rating": 4, "verified": True, |
| 5715 | "date": dates.get_date_range(19)[1], |
| 5716 | "excerpt": "Good size for the price.", |
| 5717 | "title": "Good value", |
| 5718 | }, |
| 5719 | { |
| 5720 | "score": 0, "rating": 4, "verified": False, |
| 5721 | "date": dates.get_date_range(24)[1], |
| 5722 | "excerpt": "Does the job, nothing fancy.", |
| 5723 | "title": "Fine", |
| 5724 | }, |
| 5725 | ], |
| 5726 | }, |
| 5727 | { |
| 5728 | "asin": "B0MOCK00X2", |
| 5729 | "date": dates.get_date_range(1)[1], |
| 5730 | "name": f"{subquery.search_query} Classic | Everyday Model", |
| 5731 | "short_name": "Classic", |
| 5732 | "brand": subquery.search_query.split()[0].title() if subquery.search_query else "Example", |
| 5733 | "url": "https://www.amazon.com/dp/B0MOCK00X2", |
| 5734 | "rating": 4.7, |
| 5735 | "num_ratings": 8446, |
| 5736 | "price": 24.99, |
| 5737 | "currency": "USD", |
| 5738 | "sponsored": False, |
| 5739 | "relevance": 0.8, |
| 5740 | "why_relevant": "Mock Amazon product", |
| 5741 | "product_rating": 4.7, |
| 5742 | "product_rating_count": 8446, |
| 5743 | "star_distribution": { |
| 5744 | "one_star": 120, "two_star": 90, "three_star": 300, |
| 5745 | "four_star": 1010, "five_star": 6926, |
| 5746 | }, |
| 5747 | "top_comments": [ |
| 5748 | { |
| 5749 | "score": 12, "rating": 5, "verified": True, |
| 5750 | "date": dates.get_date_range(4)[1], |
| 5751 | "excerpt": "Third one we've bought. They last.", |
| 5752 | "title": "Repeat buyer", |
| 5753 | }, |
| 5754 | ], |
| 5755 | }, |
| 5756 | { |
| 5757 | "asin": "B0MOCK00X3", |
| 5758 | "date": dates.get_date_range(1)[1], |
| 5759 | "name": f"{subquery.search_query} Mini | New Release", |
| 5760 | "short_name": "Mini", |
| 5761 | "brand": subquery.search_query.split()[0].title() if subquery.search_query else "Example", |
| 5762 | "url": "https://www.amazon.com/dp/B0MOCK00X3", |
| 5763 | "rating": None, |
| 5764 | "num_ratings": 57, |
| 5765 | "price": 19.99, |
| 5766 | "currency": "USD", |
| 5767 | "sponsored": False, |
| 5768 | "relevance": 0.72, |
| 5769 | "why_relevant": "Mock Amazon product", |
| 5770 | }, |
| 5771 | ], |
| 5772 | "jobs": [ |
| 5773 | { |
| 5774 | "id": "J1", |
| 5775 | "title": "Founding Enterprise Solutions Engineer", |
| 5776 | "url": f"https://boards.greenhouse.io/example/jobs/{slug}-1", |
| 5777 | "description": ( |
| 5778 | f"Work with enterprise customers on SSO, SOC 2, security, " |
| 5779 | f"and procurement workflows for {subquery.search_query}." |
| 5780 | ), |
| 5781 | "department": "Sales", |
| 5782 | "location": "San Francisco, CA", |
| 5783 | "date": dates.get_date_range(4)[0], |
| 5784 | "provider": "mock", |
| 5785 | "relevance": 0.8, |
| 5786 | "why_relevant": "Mock public job posting", |
| 5787 | }, |
| 5788 | { |
| 5789 | "id": "J2", |
| 5790 | "title": "Security Platform Engineer", |
| 5791 | "url": f"https://boards.greenhouse.io/example/jobs/{slug}-2", |
| 5792 | "description": "Build enterprise security, audit, and admin workflows.", |
| 5793 | "department": "Engineering", |
| 5794 | "location": "Remote", |
| 5795 | "date": dates.get_date_range(6)[0], |
| 5796 | "provider": "mock", |
| 5797 | "relevance": 0.78, |
| 5798 | "why_relevant": "Mock public job posting", |
| 5799 | }, |
| 5800 | ], |
| 5801 | } |
| 5802 | if source == "grounding": |
| 5803 | return payloads.get(source, []), { |
| 5804 | "label": subquery.label, |
| 5805 | "mock": True, |
| 5806 | "webSearchQueries": [subquery.search_query], |
| 5807 | "resultCount": 1, |
| 5808 | } |
| 5809 | return payloads.get(source, []), {} |
| 5810 |