| 1 | import Anthropic from '@anthropic-ai/sdk' |
| 2 | import { RawMessageStreamEvent } from '@anthropic-ai/sdk/resources' |
| 3 | import { GenerateContentResponse, GenerateContentResponseUsageMetadata } from '@google/genai' |
| 4 | import { BaseMessage, ChatMessage } from '@langchain/core/messages' |
| 5 | import { OpenAIClient } from '@langchain/openai' |
| 6 | import { Injectable, Logger, Optional } from '@nestjs/common' |
| 7 | import { AssetsService } from '@yikart/assets' |
| 8 | import { AppException, getErrorMessage, getErrorStack, ResponseCode, UserType } from '@yikart/common' |
| 9 | import { AiLogChannel, AiLogRepository, AiLogStatus, AiLogType, AssetType } from '@yikart/mongodb' |
| 10 | import OpenAI from 'openai' |
| 11 | import { from, merge, Observable } from 'rxjs' |
| 12 | import { catchError, concatMap, ignoreElements, last, share } from 'rxjs/operators' |
| 13 | import { config } from '../../../config' |
| 14 | import { AiAvailabilityService } from '../../ai-availability' |
| 15 | import { GeminiService } from '../libs/gemini/gemini.service' |
| 16 | import { OpenaiService } from '../libs/openai' |
| 17 | import { ModelsConfigService } from '../models-config' |
| 18 | import { RelayMediaResolverService } from '../relay-media' |
| 19 | import { |
| 20 | ChatCompletionDto, |
| 21 | ChatModelsQueryDto, |
| 22 | ChatStreamProxyDto, |
| 23 | UserChatCompletionDto, |
| 24 | UserClaudeChatProxyDto, |
| 25 | UserGeminiGenerateContentDto, |
| 26 | } from './chat.dto' |
| 27 | |
| 28 | type DeepSeekChatCompletionUsage = NonNullable<OpenAI.Chat.ChatCompletionChunk['usage']> & { |
| 29 | prompt_cache_hit_tokens?: number |
| 30 | prompt_cache_miss_tokens?: number |
| 31 | } |
| 32 | |
| 33 | interface TokenUsageDetails { |
| 34 | text?: number |
| 35 | image?: number |
| 36 | audio?: number |
| 37 | video?: number |
| 38 | cache_read?: number |
| 39 | cache_creation_5m?: number |
| 40 | cache_creation_1h?: number |
| 41 | } |
| 42 | |
| 43 | @Injectable() |
| 44 | export class ChatService { |
| 45 | private readonly logger = new Logger(ChatService.name) |
| 46 | |
| 47 | private readonly anthropic = new Anthropic({ |
| 48 | apiKey: config.ai.anthropic.apiKey, |
| 49 | baseURL: config.ai.anthropic.baseUrl, |
| 50 | }) |
| 51 | |
| 52 | constructor( |
| 53 | private readonly openaiService: OpenaiService, |
| 54 | private readonly aiLogRepo: AiLogRepository, |
| 55 | private readonly modelsConfigService: ModelsConfigService, |
| 56 | private readonly assetsService: AssetsService, |
| 57 | private readonly geminiService: GeminiService, |
| 58 | private readonly aiAvailability: AiAvailabilityService, |
| 59 | @Optional() private readonly relayMediaResolver?: RelayMediaResolverService, |
| 60 | ) {} |
| 61 | |
| 62 | /** |
| 63 | * 处理 content 中的 base64 图片，上传并替换 URL |
| 64 | */ |
| 65 | private async processBase64Images(content: string | unknown[], model: string, userId: string): Promise<string | unknown[]> { |
| 66 | if (typeof content === 'string') { |
| 67 | const base64ImageRegex = /data:image\/(png|jpeg|jpg|gif|webp);base64,([A-Za-z0-9+/=]+)/g |
| 68 | const matches = Array.from(content.matchAll(base64ImageRegex)) |
| 69 | |
| 70 | if (matches.length > 0) { |
| 71 | let processedContent = content |
| 72 | for (let i = matches.length - 1; i >= 0; i--) { |
| 73 | const match = matches[i] |
| 74 | const fullMatch = match[0] |
| 75 | const matchIndex = match.index! |
| 76 | const imageTypeName = match[1] |
| 77 | const imageData = match[2] |
| 78 | const mimeType = imageTypeName === 'jpg' ? 'jpeg' : imageTypeName |
| 79 | const fullMimeType = `image/${mimeType}` |
| 80 | const buffer = Buffer.from(imageData, 'base64') |
| 81 | const result = await this.assetsService.uploadFromBuffer(userId, buffer, { |
| 82 | type: AssetType.AiChatImage, |
| 83 | mimeType: fullMimeType, |
| 84 | }, model) |
| 85 | const url = this.assetsService.buildUrl(result.asset.path) |
| 86 | |
| 87 | const before = processedContent.substring(0, matchIndex) |
| 88 | const after = processedContent.substring(matchIndex + fullMatch.length) |
| 89 | processedContent = `${before}${url}${after}` |
| 90 | } |
| 91 | return processedContent |
| 92 | } |
| 93 | } |
| 94 | return content |
| 95 | } |
| 96 | |
| 97 | /** |
| 98 | * 处理 AIMessageChunk 的 content 中的 base64 图片 |
| 99 | */ |
| 100 | private async processAIMessageChunkContent( |
| 101 | content: string | unknown[] | undefined, |
| 102 | model: string, |
| 103 | userId: string, |
| 104 | ): Promise<string | unknown[] | undefined> { |
| 105 | if (!content) { |
| 106 | return content |
| 107 | } |
| 108 | if (typeof content === 'string') { |
| 109 | return await this.processBase64Images(content, model, userId) as string |
| 110 | } |
| 111 | if (Array.isArray(content)) { |
| 112 | return await Promise.all( |
| 113 | content.map(async (item) => { |
| 114 | if (typeof item === 'object' && item !== null && 'text' in item && typeof item.text === 'string') { |
| 115 | return { |
| 116 | ...item, |
| 117 | text: await this.processBase64Images(item.text, model, userId) as string, |
| 118 | } |
| 119 | } |
| 120 | return item |
| 121 | }), |
| 122 | ) |
| 123 | } |
| 124 | return content |
| 125 | } |
| 126 | |
| 127 | async chatCompletion(request: ChatCompletionDto, userId: string) { |
| 128 | const { messages, model, ...params } = request |
| 129 | |
| 130 | const langchainMessages: BaseMessage[] = messages.map((message) => { |
| 131 | return new ChatMessage(message) |
| 132 | }) |
| 133 | |
| 134 | const result = await this.openaiService.createChatCompletion({ |
| 135 | model, |
| 136 | messages: langchainMessages, |
| 137 | ...params, |
| 138 | modalities: params.modalities as OpenAIClient.Chat.ChatCompletionModality[], |
| 139 | }) |
| 140 | |
| 141 | const usage = result.usage_metadata |
| 142 | if (!usage) { |
| 143 | throw new AppException(ResponseCode.AiCallFailed, { error: 'Missing usage metadata' }) |
| 144 | } |
| 145 | |
| 146 | // 处理返回的 content 中的 base64 图片 |
| 147 | result.content = await this.processAIMessageChunkContent(result.content, model, userId) as typeof result.content |
| 148 | |
| 149 | return { |
| 150 | model, |
| 151 | usage, |
| 152 | ...result, |
| 153 | } |
| 154 | } |
| 155 | |
| 156 | private async handleCompletion( |
| 157 | params: ChatCompletionDto, |
| 158 | userId: string, |
| 159 | userType: UserType, |
| 160 | modelConfig: { name: string, channel: AiLogChannel }, |
| 161 | startedAt: Date, |
| 162 | usage: { input_tokens?: number, output_tokens?: number, total_tokens?: number, input_token_details?: TokenUsageDetails, output_token_details?: TokenUsageDetails }, |
| 163 | result: { model: string, usage: typeof usage }, |
| 164 | ): Promise<void> { |
| 165 | this.logger.debug({ |
| 166 | usage, |
| 167 | modelConfig, |
| 168 | }) |
| 169 | |
| 170 | const duration = Date.now() - startedAt.getTime() |
| 171 | |
| 172 | await this.aiLogRepo.create({ |
| 173 | userId, |
| 174 | userType, |
| 175 | model: params.model, |
| 176 | channel: modelConfig.channel, |
| 177 | startedAt, |
| 178 | duration, |
| 179 | type: AiLogType.Chat, |
| 180 | request: params as unknown as Record<string, unknown>, |
| 181 | response: result, |
| 182 | status: AiLogStatus.Success, |
| 183 | }) |
| 184 | } |
| 185 | |
| 186 | async userChatCompletion({ userId, userType, ...params }: UserChatCompletionDto) { |
| 187 | const modelConfig = (await this.getChatModelConfig({ userId, userType })).find((m: { name: string }) => m.name === params.model) |
| 188 | if (!modelConfig) { |
| 189 | throw new AppException(ResponseCode.InvalidModel) |
| 190 | } |
| 191 | |
| 192 | const startedAt = new Date() |
| 193 | |
| 194 | const result = await this.chatCompletion(params, userId) |
| 195 | |
| 196 | const { usage } = result |
| 197 | |
| 198 | await this.handleCompletion( |
| 199 | params, |
| 200 | userId, |
| 201 | userType, |
| 202 | modelConfig, |
| 203 | startedAt, |
| 204 | usage, |
| 205 | result, |
| 206 | ) |
| 207 | |
| 208 | return result |
| 209 | } |
| 210 | |
| 211 | /** |
| 212 | * 获取聊天模型参数 |
| 213 | * @param data 查询参数，包含可选的 userId 和 userType，可用于后续个性化模型推荐 |
| 214 | */ |
| 215 | async getChatModelConfig(data: ChatModelsQueryDto) { |
| 216 | let models = this.modelsConfigService.config.chat |
| 217 | |
| 218 | if (data.channel) { |
| 219 | models = models.filter(m => m.channel === data.channel) |
| 220 | } |
| 221 | |
| 222 | if (data.scene) { |
| 223 | models = models.filter(m => m.scenes?.includes(data.scene!)) |
| 224 | } |
| 225 | |
| 226 | return models |
| 227 | } |
| 228 | |
| 229 | private async processChunkContent( |
| 230 | chunk: OpenAI.Chat.ChatCompletionChunk, |
| 231 | model: string, |
| 232 | userId: string, |
| 233 | ): Promise<OpenAI.Chat.ChatCompletionChunk> { |
| 234 | const choice = chunk.choices[0] |
| 235 | if (!choice?.delta?.content) { |
| 236 | return chunk |
| 237 | } |
| 238 | |
| 239 | const content = choice.delta.content |
| 240 | const processedContent = await this.processBase64Images(content, model, userId) |
| 241 | if (processedContent !== content) { |
| 242 | return { |
| 243 | ...chunk, |
| 244 | choices: [{ |
| 245 | ...choice, |
| 246 | delta: { |
| 247 | ...choice.delta, |
| 248 | content: processedContent as string, |
| 249 | }, |
| 250 | }], |
| 251 | } |
| 252 | } |
| 253 | return chunk |
| 254 | } |
| 255 | |
| 256 | async proxyChatStream( |
| 257 | params: ChatStreamProxyDto & { userId: string, userType: UserType }, |
| 258 | ): Promise<Observable<OpenAI.Chat.ChatCompletionChunk>> { |
| 259 | const { userId, userType, model, ...body } = params |
| 260 | |
| 261 | const modelConfig = (await this.getChatModelConfig({ userId, userType })) |
| 262 | .find(m => m.name === model) |
| 263 | if (!modelConfig) { |
| 264 | throw new AppException(ResponseCode.InvalidModel) |
| 265 | } |
| 266 | |
| 267 | const startedAt = new Date() |
| 268 | |
| 269 | const stream = await this.openaiService.createRawStream({ |
| 270 | ...body, |
| 271 | model, |
| 272 | stream: true, |
| 273 | stream_options: { include_usage: true }, |
| 274 | } as OpenAI.Chat.ChatCompletionCreateParamsStreaming) |
| 275 | |
| 276 | const stream$ = from(stream as AsyncIterable<OpenAI.Chat.ChatCompletionChunk>).pipe(share()) |
| 277 | |
| 278 | const contentStream$ = stream$.pipe( |
| 279 | concatMap(chunk => this.processChunkContent(chunk, model, userId)), |
| 280 | ) |
| 281 | |
| 282 | const billingStream$ = stream$.pipe( |
| 283 | last(), |
| 284 | concatMap(async (lastChunk) => { |
| 285 | if (lastChunk.usage) { |
| 286 | const usage: DeepSeekChatCompletionUsage = lastChunk.usage |
| 287 | const cacheReadTokens = usage.prompt_cache_hit_tokens ?? usage.prompt_tokens_details?.cached_tokens ?? 0 |
| 288 | const inputTokens = usage.prompt_cache_miss_tokens ?? Math.max((usage.prompt_tokens || 0) - cacheReadTokens, 0) |
| 289 | const finalUsage = { |
| 290 | input_tokens: inputTokens, |
| 291 | output_tokens: usage.completion_tokens, |
| 292 | total_tokens: usage.total_tokens, |
| 293 | input_token_details: { cache_read: cacheReadTokens }, |
| 294 | } |
| 295 | await this.handleCompletion( |
| 296 | { model } as ChatCompletionDto, |
| 297 | userId, |
| 298 | userType, |
| 299 | modelConfig, |
| 300 | startedAt, |
| 301 | finalUsage, |
| 302 | { model, usage: finalUsage }, |
| 303 | ) |
| 304 | } |
| 305 | |
| 306 | await this.aiAvailability.recordSuccess( |
| 307 | { provider: 'openai', operation: 'proxyChatStream', model }, |
| 308 | Date.now() - startedAt.getTime(), |
| 309 | ) |
| 310 | }), |
| 311 | ignoreElements(), |
| 312 | ) |
| 313 | |
| 314 | return merge(contentStream$, billingStream$).pipe( |
| 315 | catchError((error) => { |
| 316 | void this.aiAvailability.recordFailure( |
| 317 | { provider: 'openai', operation: 'proxyChatStream', model }, |
| 318 | error, |
| 319 | Date.now() - startedAt.getTime(), |
| 320 | ) |
| 321 | this.logger.error(`Proxy stream error: ${getErrorMessage(error)}`) |
| 322 | throw error |
| 323 | }), |
| 324 | ) |
| 325 | } |
| 326 | |
| 327 | /** |
| 328 | * Claude 流式对话（透传） |
| 329 | * 返回 Observable<RawMessageStreamEvent> 原始事件流 |
| 330 | */ |
| 331 | async proxyClaudeChatStream({ userId, userType, ...params }: UserClaudeChatProxyDto): Promise<Observable<RawMessageStreamEvent>> { |
| 332 | const body = params |
| 333 | const modelConfig = (await this.getChatModelConfig({ userId, userType })).find((m: { name: string }) => m.name === params.model) |
| 334 | if (!modelConfig) { |
| 335 | throw new AppException(ResponseCode.InvalidModel) |
| 336 | } |
| 337 | |
| 338 | const startedAt = new Date() |
| 339 | |
| 340 | const resolvedBody = await this.resolveRelayJson(body) |
| 341 | const stream = this.anthropic.messages.stream(resolvedBody as Anthropic.MessageStreamParams) |
| 342 | |
| 343 | const stream$ = from(stream).pipe( |
| 344 | share(), |
| 345 | ) |
| 346 | |
| 347 | const contentStream$ = stream$ |
| 348 | |
| 349 | const completeStream$ = stream$.pipe( |
| 350 | last(), |
| 351 | concatMap(async () => { |
| 352 | const finalMessage = await stream.finalMessage() |
| 353 | const usage = finalMessage.usage |
| 354 | const cacheCreation = usage.cache_creation |
| 355 | const inputTokenDetails: TokenUsageDetails = { |
| 356 | cache_read: usage.cache_read_input_tokens ?? 0, |
| 357 | cache_creation_5m: cacheCreation?.ephemeral_5m_input_tokens ?? usage.cache_creation_input_tokens ?? 0, |
| 358 | cache_creation_1h: cacheCreation?.ephemeral_1h_input_tokens ?? 0, |
| 359 | } |
| 360 | const finalUsage = { |
| 361 | input_tokens: usage.input_tokens, |
| 362 | output_tokens: usage.output_tokens, |
| 363 | input_token_details: inputTokenDetails, |
| 364 | } |
| 365 | |
| 366 | await this.handleCompletion( |
| 367 | { model: params.model, messages: params.messages as ChatCompletionDto['messages'] }, |
| 368 | userId, |
| 369 | userType, |
| 370 | modelConfig, |
| 371 | startedAt, |
| 372 | finalUsage, |
| 373 | { model: params.model, usage: finalUsage }, |
| 374 | ) |
| 375 | |
| 376 | await this.aiAvailability.recordSuccess( |
| 377 | { provider: 'anthropic', operation: 'proxyClaudeChatStream', model: params.model }, |
| 378 | Date.now() - startedAt.getTime(), |
| 379 | ) |
| 380 | }), |
| 381 | ignoreElements(), |
| 382 | ) |
| 383 | |
| 384 | return merge(contentStream$, completeStream$).pipe( |
| 385 | catchError((error) => { |
| 386 | void this.aiAvailability.recordFailure( |
| 387 | { provider: 'anthropic', operation: 'proxyClaudeChatStream', model: params.model }, |
| 388 | error, |
| 389 | Date.now() - startedAt.getTime(), |
| 390 | ) |
| 391 | this.logger.error(`Error in proxyClaudeChatStream: ${getErrorMessage(error)}`, getErrorStack(error)) |
| 392 | throw error |
| 393 | }), |
| 394 | ) |
| 395 | } |
| 396 | |
| 397 | /** |
| 398 | * Gemini generateContent（通用内容生成，支持视频/音频/图片分析） |
| 399 | */ |
| 400 | async userGeminiGenerateContent(request: UserGeminiGenerateContentDto) { |
| 401 | const { userId, userType, model, ...params } = request |
| 402 | |
| 403 | // 获取模型配置 |
| 404 | const modelConfig = (await this.getChatModelConfig({ userId, userType })) |
| 405 | .find(m => m.name === model) |
| 406 | if (!modelConfig) { |
| 407 | throw new AppException(ResponseCode.InvalidModel) |
| 408 | } |
| 409 | |
| 410 | const startedAt = new Date() |
| 411 | |
| 412 | let result: GenerateContentResponse | undefined |
| 413 | let usage: GenerateContentResponseUsageMetadata | undefined |
| 414 | |
| 415 | try { |
| 416 | // 调用 Gemini generateContent |
| 417 | const responses = await this.geminiService.generateContentStream({ |
| 418 | model, |
| 419 | contents: params.contents, |
| 420 | config: params.config, |
| 421 | }) |
| 422 | |
| 423 | for await (const chunk of responses) { |
| 424 | usage = chunk.usageMetadata |
| 425 | |
| 426 | if (!result) { |
| 427 | result = chunk |
| 428 | } |
| 429 | else { |
| 430 | const existingParts = result.candidates?.[0]?.content?.parts || [] |
| 431 | const newParts = chunk.candidates?.[0]?.content?.parts || [] |
| 432 | if (result.candidates?.[0]?.content) { |
| 433 | result.candidates[0].content.parts = [...existingParts, ...newParts] |
| 434 | } |
| 435 | result.usageMetadata = chunk.usageMetadata |
| 436 | } |
| 437 | } |
| 438 | |
| 439 | await this.aiAvailability.recordSuccess( |
| 440 | { provider: 'gemini', operation: 'generateContentStream', model }, |
| 441 | Date.now() - startedAt.getTime(), |
| 442 | ) |
| 443 | } |
| 444 | catch (error) { |
| 445 | await this.aiAvailability.recordFailure( |
| 446 | { provider: 'gemini', operation: 'generateContentStream', model }, |
| 447 | error, |
| 448 | Date.now() - startedAt.getTime(), |
| 449 | ) |
| 450 | throw error |
| 451 | } |
| 452 | |
| 453 | const finalUsage = this.buildGeminiUsage(usage) |
| 454 | |
| 455 | await this.handleCompletion( |
| 456 | { model, messages: [] }, |
| 457 | userId, |
| 458 | userType, |
| 459 | modelConfig, |
| 460 | startedAt, |
| 461 | finalUsage, |
| 462 | { model, usage: finalUsage }, |
| 463 | ) |
| 464 | |
| 465 | return result! |
| 466 | } |
| 467 | |
| 468 | /** |
| 469 | * 将 Gemini usageMetadata 转换为统一 usage 结构。 |
| 470 | */ |
| 471 | buildGeminiUsage(usage: GenerateContentResponseUsageMetadata | undefined): { |
| 472 | input_tokens: number |
| 473 | output_tokens: number |
| 474 | total_tokens: number |
| 475 | input_token_details: TokenUsageDetails |
| 476 | output_token_details: TokenUsageDetails | undefined |
| 477 | } { |
| 478 | const inputTokenDetails = this.extractGeminiTokenDetails(usage?.promptTokensDetails) ?? {} |
| 479 | const outputTokenDetails = this.extractGeminiTokenDetails(usage?.candidatesTokensDetails) |
| 480 | const cacheReadTokens = usage?.cachedContentTokenCount || 0 |
| 481 | if (inputTokenDetails.text) { |
| 482 | inputTokenDetails.text = Math.max(inputTokenDetails.text - cacheReadTokens, 0) |
| 483 | } |
| 484 | inputTokenDetails.cache_read = cacheReadTokens |
| 485 | |
| 486 | return { |
| 487 | input_tokens: cacheReadTokens > 0 ? Math.max((usage?.promptTokenCount || 0) - cacheReadTokens, 0) : usage?.promptTokenCount || 0, |
| 488 | output_tokens: usage?.candidatesTokenCount || 0, |
| 489 | total_tokens: usage?.totalTokenCount || 0, |
| 490 | input_token_details: inputTokenDetails, |
| 491 | output_token_details: outputTokenDetails, |
| 492 | } |
| 493 | } |
| 494 | |
| 495 | private async resolveRelayJson<T>(value: T): Promise<T> { |
| 496 | if (!this.relayMediaResolver) { |
| 497 | return value |
| 498 | } |
| 499 | return await this.relayMediaResolver.resolveJson(value) |
| 500 | } |
| 501 | |
| 502 | extractGeminiTokenDetails(details: GenerateContentResponseUsageMetadata['promptTokensDetails']): TokenUsageDetails | undefined { |
| 503 | if (!details) { |
| 504 | return undefined |
| 505 | } |
| 506 | |
| 507 | const result: TokenUsageDetails = {} |
| 508 | for (const detail of details) { |
| 509 | const rawModality = detail.modality |
| 510 | const rawTokenCount = detail.tokenCount |
| 511 | if (typeof rawModality !== 'string' || typeof rawTokenCount !== 'number' || rawTokenCount <= 0) { |
| 512 | continue |
| 513 | } |
| 514 | |
| 515 | const modality = rawModality.toLowerCase() |
| 516 | if (modality.includes('text')) { |
| 517 | result.text = (result.text || 0) + rawTokenCount |
| 518 | } |
| 519 | else if (modality.includes('image')) { |
| 520 | result.image = (result.image || 0) + rawTokenCount |
| 521 | } |
| 522 | else if (modality.includes('audio')) { |
| 523 | result.audio = (result.audio || 0) + rawTokenCount |
| 524 | } |
| 525 | else if (modality.includes('video')) { |
| 526 | result.video = (result.video || 0) + rawTokenCount |
| 527 | } |
| 528 | } |
| 529 | |
| 530 | return Object.keys(result).length > 0 ? result : undefined |
| 531 | } |
| 532 | } |
| 533 |