| 1 | import { createRequire } from 'module' |
| 2 | import fs from 'fs' |
| 3 | import path from 'path' |
| 4 | import log from 'electron-log/main.js' |
| 5 | import { nanoid } from 'nanoid' |
| 6 | import { HumanMessage } from '@langchain/core/messages' |
| 7 | import { resolveModelTimeoutMs } from '@shared/model-timeout' |
| 8 | import { isSupportedImageMimeType, normalizeImageMimeType } from '@shared/image-mime' |
| 9 | import { resolveModel } from '../agent' |
| 10 | import type { ModelRuntimeConfig } from '../agent' |
| 11 | import { extractModelText } from '../ipc/utils' |
| 12 | import type { ThinkingSource } from '@shared/thinking' |
| 13 | |
| 14 | const require = createRequire(import.meta.url) |
| 15 | const mammoth = require('mammoth') as typeof import('mammoth') |
| 16 | const TurndownService = require('turndown') as new (options?: Record<string, unknown>) => { |
| 17 | use: (plugin: unknown) => void |
| 18 | turndown: (html: string) => string |
| 19 | } |
| 20 | const { gfm } = require('@joplin/turndown-plugin-gfm') as { gfm: unknown } |
| 21 | |
| 22 | const NULL_CHAR_PATTERN = new RegExp(String.fromCharCode(0), 'g') |
| 23 | |
| 24 | const SUPPORTED_EXTENSIONS = new Set(['.md', '.txt', '.text', '.csv', '.docx']) |
| 25 | const SUPPORTED_IMAGE_EXTENSIONS = new Set(['.png', '.jpg', '.jpeg', '.webp']) |
| 26 | const MAX_FILE_SIZE = 10 * 1024 * 1024 |
| 27 | const MAX_IMAGE_SIZE = 5 * 1024 * 1024 |
| 28 | |
| 29 | const stripControlChars = (value: string): string => |
| 30 | value.replace(NULL_CHAR_PATTERN, '').replace(/\r\n/g, '\n').replace(/\r/g, '\n') |
| 31 | |
| 32 | const compactText = (value: string): string => |
| 33 | stripControlChars(value) |
| 34 | .split('\n') |
| 35 | .map((line) => line.replace(/[ \t]+/g, ' ').trim()) |
| 36 | .join('\n') |
| 37 | .replace(/\n{4,}/g, '\n\n\n') |
| 38 | .trim() |
| 39 | |
| 40 | const stripInlineImagesFromHtml = (html: string): string => |
| 41 | html.replace(/<img\b[^>]*>/gi, (tag) => { |
| 42 | const alt = tag.match(/\balt=(["'])(.*?)\1/i)?.[2]?.trim() |
| 43 | return alt ? `<p>[图片：${alt}]</p>` : '' |
| 44 | }) |
| 45 | |
| 46 | const stripMarkdownDataImages = (markdown: string): string => |
| 47 | markdown.replace(/!\[[^\]]*]\(data:[^)]+\)/gi, '').replace(/!\[[^\]]*]\(\s*\)/g, '') |
| 48 | |
| 49 | const convertDocxToMarkdown = async (filePath: string): Promise<string> => { |
| 50 | const result = await mammoth.convertToHtml({ path: filePath }) |
| 51 | const turndown = new TurndownService({ |
| 52 | headingStyle: 'atx', |
| 53 | bulletListMarker: '-', |
| 54 | codeBlockStyle: 'fenced' |
| 55 | }) |
| 56 | turndown.use(gfm) |
| 57 | return compactText( |
| 58 | stripMarkdownDataImages(turndown.turndown(stripInlineImagesFromHtml(result.value))) |
| 59 | ) |
| 60 | } |
| 61 | |
| 62 | const toSafeFileName = (value: string): string => |
| 63 | value |
| 64 | .replace(/[\\/:"*?<>|]+/g, '-') |
| 65 | .replace(/\s+/g, '-') |
| 66 | .replace(/^-+|-+$/g, '') |
| 67 | .slice(0, 80) || 'source' |
| 68 | |
| 69 | const mimeTypeFromExtension = (ext: string): string => { |
| 70 | if (ext === '.png') return 'image/png' |
| 71 | if (ext === '.jpg' || ext === '.jpeg') return 'image/jpeg' |
| 72 | if (ext === '.webp') return 'image/webp' |
| 73 | return '' |
| 74 | } |
| 75 | |
| 76 | const buildImageTextExtractionPrompt = (name: string): string => |
| 77 | [ |
| 78 | `请只识别这张图片中的文字、数字、表格、图表标签、界面文案或可直接引用的数据。图片文件名：${name}`, |
| 79 | '', |
| 80 | '严格要求：', |
| 81 | '- 只做文字/数据识别，不做视觉理解、图片摘要、用途建议或风格分析。', |
| 82 | '- 不要描述配色、构图、审美、质感、版式、插画风格或设计方向。', |
| 83 | '- 不要判断这张图适合做封面、插图、风格参考或证据素材。', |
| 84 | '- 不要编造图片中不存在的文字、数字或事实。', |
| 85 | '- 如果没有可识别文字或数据，只输出“未识别到明确文字内容”。', |
| 86 | '', |
| 87 | '请按以下 Markdown 小节输出：', |
| 88 | '## Extracted Text', |
| 89 | '## Extracted Data' |
| 90 | ].join('\n') |
| 91 | |
| 92 | const buildFallbackImageTextExtraction = (name: string): string => |
| 93 | [ |
| 94 | '## Extracted Text', |
| 95 | `图片 ${name} 已上传，但未识别到明确文字内容。`, |
| 96 | '', |
| 97 | '## Extracted Data', |
| 98 | '- 未识别到明确数据。' |
| 99 | ].join('\n') |
| 100 | |
| 101 | export type SourceKind = ThinkingSource['kind'] |
| 102 | |
| 103 | export interface PreparedSource { |
| 104 | id: string |
| 105 | name: string |
| 106 | kind: SourceKind |
| 107 | sourcePath: string |
| 108 | assetsPath?: string |
| 109 | } |
| 110 | |
| 111 | export interface ImageTextExtractionOptions { |
| 112 | provider: string |
| 113 | apiKey: string |
| 114 | model: string |
| 115 | baseUrl: string |
| 116 | maxTokens?: number |
| 117 | modelRuntime?: ModelRuntimeConfig |
| 118 | modelTimeoutMs: number |
| 119 | } |
| 120 | |
| 121 | const IMAGE_TEXT_MARKER = '<!-- image-text-extracted -->' |
| 122 | |
| 123 | function detectKind(ext: string): SourceKind { |
| 124 | if (SUPPORTED_IMAGE_EXTENSIONS.has(ext)) return 'image' |
| 125 | if (ext === '.docx') return 'docx' |
| 126 | if (ext === '.md') return 'markdown' |
| 127 | if (ext === '.csv') return 'csv' |
| 128 | return 'text' |
| 129 | } |
| 130 | |
| 131 | async function extractImageText(args: { |
| 132 | filePath: string |
| 133 | name: string |
| 134 | ext: string |
| 135 | options: ImageTextExtractionOptions |
| 136 | }): Promise<string> { |
| 137 | const mimeType = normalizeImageMimeType(mimeTypeFromExtension(args.ext)) |
| 138 | if (!isSupportedImageMimeType(mimeType)) { |
| 139 | throw new Error(`不支持的图片格式：${mimeType || 'unknown'}`) |
| 140 | } |
| 141 | const imageBase64 = await fs.promises.readFile(args.filePath, 'base64') |
| 142 | const imageBytes = Buffer.byteLength(imageBase64, 'base64') |
| 143 | log.info('[thinking:image-text] invoke thinking image text model', { |
| 144 | provider: args.options.provider, |
| 145 | model: args.options.model, |
| 146 | mimeType, |
| 147 | imageBytes |
| 148 | }) |
| 149 | const model = resolveModel( |
| 150 | args.options.provider, |
| 151 | args.options.apiKey, |
| 152 | args.options.model, |
| 153 | args.options.baseUrl, |
| 154 | 0.1, |
| 155 | args.options.maxTokens, |
| 156 | args.options.modelRuntime |
| 157 | ) |
| 158 | const result = await model.invoke( |
| 159 | [ |
| 160 | new HumanMessage({ |
| 161 | content: [ |
| 162 | { type: 'text', text: buildImageTextExtractionPrompt(args.name) }, |
| 163 | { type: 'image_url', image_url: { url: `data:${mimeType};base64,${imageBase64}` } } |
| 164 | ] |
| 165 | }) |
| 166 | ], |
| 167 | { |
| 168 | signal: AbortSignal.timeout(resolveModelTimeoutMs(args.options.modelTimeoutMs, 'document')) |
| 169 | } |
| 170 | ) |
| 171 | const response = extractModelText(result) |
| 172 | return response.trim() || buildFallbackImageTextExtraction(args.name) |
| 173 | } |
| 174 | |
| 175 | const parseImageAssetPath = (content: string): string => { |
| 176 | const match = content.match(/^- thinkingAssetPath:\s*(.+)$/m) |
| 177 | return match?.[1]?.trim() || '' |
| 178 | } |
| 179 | |
| 180 | export async function extractPendingImageTextSources( |
| 181 | thinkingDir: string, |
| 182 | options: ImageTextExtractionOptions |
| 183 | ): Promise<void> { |
| 184 | const sourcesDir = path.join(thinkingDir, 'sources') |
| 185 | if (!fs.existsSync(sourcesDir)) return |
| 186 | |
| 187 | const entries = await fs.promises.readdir(sourcesDir, { withFileTypes: true }) |
| 188 | for (const entry of entries) { |
| 189 | if (!entry.isFile() || !entry.name.endsWith('.image.md')) continue |
| 190 | const sourcePath = path.join(sourcesDir, entry.name) |
| 191 | const content = await fs.promises.readFile(sourcePath, 'utf-8') |
| 192 | if (content.includes(IMAGE_TEXT_MARKER)) continue |
| 193 | |
| 194 | const imagePath = parseImageAssetPath(content) |
| 195 | if (!imagePath || !fs.existsSync(imagePath)) continue |
| 196 | |
| 197 | try { |
| 198 | const extractedText = await extractImageText({ |
| 199 | filePath: imagePath, |
| 200 | name: path.basename(imagePath), |
| 201 | ext: path.extname(imagePath).toLowerCase(), |
| 202 | options |
| 203 | }) |
| 204 | await fs.promises.writeFile( |
| 205 | sourcePath, |
| 206 | [ |
| 207 | content.trimEnd(), |
| 208 | '', |
| 209 | IMAGE_TEXT_MARKER, |
| 210 | '## Notes', |
| 211 | '- 以下内容仅来自图片中的文字/数据识别，不包含风格、配色、构图或用途分析。', |
| 212 | '', |
| 213 | extractedText.trim() |
| 214 | ].join('\n') + '\n', |
| 215 | 'utf-8' |
| 216 | ) |
| 217 | } catch (err) { |
| 218 | log.warn('[thinking:source-prepare] image text extraction failed', { |
| 219 | source: entry.name, |
| 220 | message: err instanceof Error ? err.message : String(err) |
| 221 | }) |
| 222 | await fs.promises.writeFile( |
| 223 | sourcePath, |
| 224 | [ |
| 225 | content.trimEnd(), |
| 226 | '', |
| 227 | IMAGE_TEXT_MARKER, |
| 228 | '## Extracted Text', |
| 229 | '- 图片文字识别失败。', |
| 230 | '', |
| 231 | '## Extracted Data', |
| 232 | '- 未识别到明确数据。' |
| 233 | ].join('\n') + '\n', |
| 234 | 'utf-8' |
| 235 | ) |
| 236 | } |
| 237 | } |
| 238 | } |
| 239 | |
| 240 | export async function prepareSourceFile( |
| 241 | filePath: string, |
| 242 | thinkingDir: string |
| 243 | ): Promise<PreparedSource> { |
| 244 | const resolved = path.resolve(filePath) |
| 245 | const stat = await fs.promises.stat(resolved) |
| 246 | if (!stat.isFile()) throw new Error(`Not a file: ${resolved}`) |
| 247 | |
| 248 | const ext = path.extname(resolved).toLowerCase() |
| 249 | const isImage = SUPPORTED_IMAGE_EXTENSIONS.has(ext) |
| 250 | |
| 251 | if (!SUPPORTED_EXTENSIONS.has(ext) && !isImage) { |
| 252 | throw new Error(`Unsupported file type: ${ext}`) |
| 253 | } |
| 254 | |
| 255 | if (isImage && stat.size > MAX_IMAGE_SIZE) { |
| 256 | throw new Error(`Image file too large (max 5MB): ${path.basename(resolved)}`) |
| 257 | } |
| 258 | if (!isImage && stat.size > MAX_FILE_SIZE) { |
| 259 | throw new Error(`File too large (max 10MB): ${path.basename(resolved)}`) |
| 260 | } |
| 261 | |
| 262 | const sourcesDir = path.join(thinkingDir, 'sources') |
| 263 | const assetsDir = path.join(thinkingDir, 'assets') |
| 264 | await fs.promises.mkdir(sourcesDir, { recursive: true }) |
| 265 | await fs.promises.mkdir(assetsDir, { recursive: true }) |
| 266 | |
| 267 | const kind = detectKind(ext) |
| 268 | const baseName = path.basename(resolved, ext) |
| 269 | const safeName = toSafeFileName(baseName) |
| 270 | const stamp = Date.now() |
| 271 | const uid = nanoid(8) |
| 272 | |
| 273 | const id = `${stamp}-${uid}-${safeName}` |
| 274 | let sourcePath: string |
| 275 | let assetsPath: string | undefined |
| 276 | |
| 277 | if (kind === 'docx') { |
| 278 | const mdName = `${id}.md` |
| 279 | sourcePath = path.join(sourcesDir, mdName) |
| 280 | const markdown = await convertDocxToMarkdown(resolved) |
| 281 | await fs.promises.writeFile( |
| 282 | sourcePath, |
| 283 | [`# ${baseName}`, '', `> Converted from Word .docx`, '', markdown].join('\n'), |
| 284 | 'utf-8' |
| 285 | ) |
| 286 | } else if (kind === 'image') { |
| 287 | const imgName = `${id}${ext}` |
| 288 | const mdName = `${id}.image.md` |
| 289 | sourcePath = path.join(sourcesDir, mdName) |
| 290 | assetsPath = path.join(assetsDir, imgName) |
| 291 | await fs.promises.copyFile(resolved, assetsPath) |
| 292 | await fs.promises.writeFile( |
| 293 | sourcePath, |
| 294 | [ |
| 295 | `# 图片：${path.basename(resolved)}`, |
| 296 | '', |
| 297 | '## Asset', |
| 298 | `- assetId: ${id}`, |
| 299 | `- fileName: ${imgName}`, |
| 300 | `- originalPath: ${resolved}`, |
| 301 | `- thinkingAssetPath: ${assetsPath}`, |
| 302 | `- thinkingPublicPath: assets/${imgName}`, |
| 303 | '- sessionAssetPath: (set during generation copy)', |
| 304 | '- publicPath: (set during generation copy)', |
| 305 | '', |
| 306 | '## Notes', |
| 307 | '- 图片已复制到素材库，上传阶段不进行识别或解析。', |
| 308 | '- 用户发送消息后才会识别图片中的文字/数据；不会解析图片风格。' |
| 309 | ].join('\n'), |
| 310 | 'utf-8' |
| 311 | ) |
| 312 | } else { |
| 313 | const fileName = `${id}${ext}` |
| 314 | sourcePath = path.join(sourcesDir, fileName) |
| 315 | await fs.promises.copyFile(resolved, sourcePath) |
| 316 | } |
| 317 | |
| 318 | log.info('[thinking:source-prepare] prepared', { |
| 319 | kind, |
| 320 | name: path.basename(resolved), |
| 321 | id |
| 322 | }) |
| 323 | |
| 324 | return { |
| 325 | id, |
| 326 | name: path.basename(resolved), |
| 327 | kind, |
| 328 | sourcePath, |
| 329 | assetsPath |
| 330 | } |
| 331 | } |
| 332 | |
| 333 | export async function prepareMultipleSources( |
| 334 | filePaths: string[], |
| 335 | thinkingDir: string |
| 336 | ): Promise<PreparedSource[]> { |
| 337 | const results: PreparedSource[] = [] |
| 338 | for (const filePath of filePaths) { |
| 339 | results.push(await prepareSourceFile(filePath, thinkingDir)) |
| 340 | } |
| 341 | return results |
| 342 | } |
| 343 |