{"repo": "1622359590/ai-cs-hub", "files": {"backend/services/rag.js": {"ff8d17b14023e5654fa002ec3900c26e454165c2": {"t": 1785746671.4737246, "n": 23110, "h": "d181ea2ca2052cde", "c": "/**\n * RAG 检索服务 — 混合检索（BM25 + 向量）+ Re-ranking\n * 向量检索使用真实 Embedding API（DeepSeek / OpenAI 兼容）\n */\nconst { getDb } = require('../database/schema');\n\nlet sqliteVecLoaded = false;\nfunction ensureVec(db) {\n  if (sqliteVecLoaded) return;\n  try {\n    const sqliteVec = require('sqlite-vec');\n    sqliteVec.load(db);\n    sqliteVecLoaded = true;\n  } catch (e) {\n    console.warn('sqlite-vec 加载失败:', e.message);\n  }\n}\n\n// Embedding 配置\nconst EMBEDDING_DIMENSION = 512; // bge-small-zh-v1.5 输出维度\nconst EMBEDDING_BATCH_SIZE = 10; // DashScope 限制每批最多 10 条\n\n// 命名空间前缀（用于 chunk ID，避免冲突）\nconst NS = { knowledge: 'k', tutorial: 't', faq: 'f' };\n\n// Embedding 模型映射\nconst EMBEDDING_MODELS = {\n  deepseek: 'deepseek-embedding',\n  openai: 'text-embedding-3-small',\n  qwen: 'text-embedding-v3',\n};\n\n// Provider 对应的 Embedding API URL\nfunction getEmbeddingUrl(provider) {\n  const urls = {\n    deepseek: 'https://api.deepseek.com/embeddings',\n    openai: 'https://api.openai.com/v1/embeddings',\n    qwen: 'https://dashscope.aliyuncs.com/compatible-mode/v1/embeddings',\n  };\n  return urls[provider] || '';\n}\n\n// BM25 参数\nconst BM25_K1 = 1.5;\nconst BM25_B = 0.75;\n\n// 同义词映射（用于查询扩展）\nconst SYNONYMS = {\n  '工单': ['工单', '工单', '问题', '反馈', '需求'],\n  '教程': ['教程', '课程', '学习', '视频'],\n  '客服': ['客服', '人工', '帮助', '支持'],\n};\n\n// 中文停用词\nconst STOP_WORDS = new Set([\n  '的', '了', '是', '在', '我', '有', '和', '就', '不', '人', '都', '一',\n  '一个', '上', '也', '很', '到', '说', '要', '去', '你', '会', '着', '没有',\n  '看', '好', '自己', '这', '他', '么', '把', '那', '你', '它', '她', '哪',\n  '如何', '怎么', '什么', '为什么', '可以', '能', '可以', '请', '求', '帮',\n]);\n\n/**\n * 分词（简单实现：中文逐字 + 英文按空格）\n */\nfunction tokenize(text) {\n  if (!text) return [];\n  const tokens = [];\n  // 匹配中文字符或英文单词\n  const matches = text.match(/[\\u4e00-\\u9fff]|[a-zA-Z0-9]+/g) || [];\n  for (const m of matches) {\n    const lower = m.toLowerCase();\n    if (!STOP_WORDS.has(lower) && lower.length > 0) {\n      tokens.push(lower);\n    }\n  }\n  return tokens;\n}\n\n/**\n * 查询扩展：添加同义词\n */\nfunction expandQuery(query) {\n  let expanded = query;\n  for (const [key, synonyms] of Object.entries(SYNONYMS)) {\n    if (query.includes(key)) {\n      expanded += ' ' + synonyms.join(' ');\n    }\n  }\n  return expanded;\n}\n\n// ===== BM25 索引 =====\nclass BM25Index {\n  constructor() {\n    this.docs = [];\n    this.docCount = 0;\n    this.avgDocLen = 0;\n    this.docFreq = {}; // term -> 文档频率\n    this.docLens = [];\n    this.invertedIndex = {}; // term -> [{docIdx, tf}]\n  }\n\n  build(items) {\n    this.docs = items;\n    this.docCount = items.length;\n    this.docFreq = {};\n    this.invertedIndex = {};\n    this.docLens = [];\n\n    let totalLen = 0;\n    for (let i = 0; i < items.length; i++) {\n      const tokens = tokenize(items[i].title + ' ' + items[i].content);\n      this.docLens.push(tokens.length);\n      totalLen += tokens.length;\n\n      // 统计词频\n      const tf = {};\n      for (const t of tokens) tf[t] = (tf[t] || 0) + 1;\n\n      // 构建倒排索引\n      for (const [term, freq] of Object.entries(tf)) {\n        if (!this.invertedIndex[term]) this.invertedIndex[term] = [];\n        this.invertedIndex[term].push({ docIdx: i, tf: freq });\n        this.docFreq[term] = (this.docFreq[term] || 0) + 1;\n      }\n    }\n    this.avgDocLen = totalLen / this.docCount || 1;\n  }\n\n  search(query, topK = 5) {\n    const expandedQuery = expandQuery(query);\n    const queryTokens = tokenize(expandedQuery);\n    if (queryTokens.length === 0) return [];\n\n    const scores = new Array(this.docCount).fill(0);\n\n    for (const term of queryTokens) {\n      const postings = this.invertedIndex[term] || [];\n      const df = this.docFreq[term] || 0;\n      if (df === 0) continue;\n\n      // IDF 部分\n      const idf = Math.log((this.docCount - df + 0.5) / (df + 0.5) + 1);\n\n      for (const { docIdx, tf } of postings) {\n        const docLen = this.docLens[docIdx];\n        // BM25 公式\n        const tfNorm = (tf * (BM25_K1 + 1)) / (tf + BM25_K1 * (1 - BM25_B + BM25_B * docLen / this.avgDocLen));\n        scores[docIdx] += idf * tfNorm;\n      }\n    }\n\n    // 排序取 Top-K\n    const results = [];\n    for (let i = 0; i < this.docCount; i++) {\n      if (scores[i] > 0) {\n        results.push({ id: this.docs[i].id, score: scores[i] });\n      }\n    }\n    results.sort((a, b) => b.score - a.score);\n    return results.slice(0, topK);\n  }\n}\n\n/**\n * 调用 Embedding API（异步，使用 fetch）\n * @param {string} text - 要嵌入的文本\n * @param {object} config - {provider, apiKey, baseUrl, model}\n * @returns {Promise<Float32Array|null>} 嵌入向量\n */\nasync function callEmbeddingAPI(text, config) {\n  const url = config.baseUrl || getEmbeddingUrl(config.provider);\n  if (!url || !config.apiKey) return null;\n\n  const model = config.model || EMBEDDING_MODELS[config.provider] || 'deepseek-embedding';\n\n  try {\n    const resp = await fetch(url, {\n      method: 'POST',\n      headers: {\n        'Content-Type': 'application/json',\n        'Authorization': `Bearer ${config.apiKey}`,\n      },\n      body: JSON.stringify({\n        model,\n        input: text.slice(0, 8000),\n      }),\n      signal: AbortSignal.timeout(30000),\n    });\n\n    if (!resp.ok) {\n      console.warn('Embedding API HTTP 错误:', resp.status);\n      return null;\n    }\n\n    const json = await resp.json();\n    if (json.data && json.data[0] && json.data[0].embedding) {\n      return new Float32Array(json.data[0].embedding);\n    }\n    if (json.error) {\n      console.warn('Embedding API 错误:', json.error.message || json.error);\n    }\n    return null;\n  } catch (e) {\n    console.warn('Embedding API 调用失败:', e.message);\n    return null;\n  }\n}\n\n/**\n * 批量调用 Embedding API（异步）\n * @param {string[]} texts - 文本数组\n * @param {object} config - API 配置\n * @returns {Promise<Float32Array[]>} 嵌入向量数组（失败的用 null 占位）\n */\nasync function callEmbeddingBatch(texts, config) {\n  const url = config.baseUrl || getEmbeddingUrl(config.provider);\n  if (!url || !config.apiKey) return texts.map(() => null);\n\n  const model = config.model || EMBEDDING_MODELS[config.provider] || 'deepseek-embedding';\n  const results = new Array(texts.length).fill(null);\n\n  // 分批处理\n  for (let i = 0; i < texts.length; i += EMBEDDING_BATCH_SIZE) {\n    const batch = texts.slice(i, i + EMBEDDING_BATCH_SIZE);\n\n    try {\n      const resp = await fetch(url, {\n        method: 'POST',\n        headers: {\n          'Content-Type': 'application/json',\n          'Authorization': `Bearer ${config.apiKey}`,\n        },\n        body: JSON.stringify({\n          model,\n          input: batch.map(t => t.slice(0, 8000)),\n        }),\n        signal: AbortSignal.timeout(60000),\n      });\n\n      if (!resp.ok) {\n        console.warn('Embedding 批量 API HTTP 错误:', resp.status);\n        continue;\n      }\n\n      const json = await resp.json();\n      if (json.data) {\n        for (const item of json.data) {\n          results[i + item.index] = new Float32Array(item.embedding);\n        }\n      }\n      if (json.error) {\n        console.warn('Embedding 批量 API 错误:', json.error.message || json.error);\n      }\n    } catch (e) {\n      console.warn('Embedding 批量 API 调用失败:', e.message);\n    }\n\n    // 批次间隔，避免限流\n    if (i + EMBEDDING_BATCH_SIZE < texts.length) {\n      await new Promise(r => setTimeout(r, 1000));\n    }\n  }\n\n  return results;\n}\n\n/**\n * 从 settings 读取 Embedding 配置\n * 支持独立的 embedding 配置，回退到通用 ai 配置\n */\nfunction getEmbeddingConfig() {\n  const db = getDb();\n  const rows = db.prepare(\n    \"SELECT key, value FROM settings WHERE key IN ('embedding_provider','embedding_api_key','embedding_base_url','embedding_model','ai_provider','ai_api_key','ai_base_url')\"\n  ).all();\n  const config = {};\n  for (const row of rows) config[row.key] = row.value;\n  return {\n    provider: config.embedding_provider || config.ai_provider || 'deepseek',\n    apiKey: config.embedding_api_key || config.ai_api_key || '',\n    baseUrl: config.embedding_base_url || '',\n    model: config.embedding_model || '',\n  };\n}\n\n// ===== 向量索引（使用 sqlite-vec）=====\n// 向量表用自增整数 ID，chunkMeta 映射 vectorId → {source, parentId, chunkIndex}\nlet vectorIdCounter = 0;\nlet chunkMeta = {}; // vectorId → {source, parentId, chunkIndex, chunkId}\nlet chunkIdToVectorId = {}; // chunkId → vectorId\n\nclass VectorIndex {\n  constructor() {\n    this.dimension = EMBEDDING_DIMENSION;\n  }\n\n  initTable(db) {\n    ensureVec(db);\n    db.exec(`\n      CREATE VIRTUAL TABLE IF NOT EXISTS knowledge_vec USING vec0(\n        id INTEGER PRIMARY KEY,\n        embedding float[${this.dimension}]\n      );\n    `);\n  }\n\n  rebuildTable(db) {\n    ensureVec(db);\n    try {\n      const info = db.prepare(\"SELECT sql FROM sqlite_master WHERE name = 'knowledge_vec'\").get();\n      if (info && !info.sql.includes(`float[${this.dimension}]`)) {\n        console.log(`向量表维度变化，重建表 (${this.dimension}维)`);\n        db.prepare('DROP TABLE IF EXISTS knowledge_vec').run();\n      }\n    } catch {}\n    this.initTable(db);\n  }\n\n  upsert(db, vectorId, embedding) {\n    if (!embedding) return false;\n    try {\n      db.prepare('DELETE FROM knowledge_vec WHERE id = ?').run(vectorId);\n      db.prepare('INSERT INTO knowledge_vec (id, embedding) VALUES (?, ?)').run(\n        vectorId,\n        Buffer.from(embedding.buffer)\n      );\n      return true;\n    } catch (e) {\n      return false;\n    }\n  }\n\n  search(db, queryVector, topK = 5) {\n    ensureVec(db);\n    if (!queryVector) return [];\n    try {\n      const results = db.prepare(`\n        SELECT id, distance FROM knowledge_vec WHERE embedding MATCH ? ORDER BY distance LIMIT ?\n      `).all(Buffer.from(queryVector.buffer), topK);\n      return results.map(r => ({ vectorId: r.id, score: 1 - r.distance }));\n    } catch (e) {\n      return [];\n    }\n  }\n}\n\n// 全局实例\nlet bm25Index = null;\nconst vectorIndex = new VectorIndex();\n\n// ===== 文档拆分 (Chunking) =====\nconst CHUNK_MAX_CHARS = 400;\nconst CHUNK_OVERLAP = 50;\n\n/**\n * 将长文本按段落拆分成小块\n */\nfunction chunkText(text, maxChars = CHUNK_MAX_CHARS) {\n  if (!text || text.length <= maxChars) return [text];\n\n  const chunks = [];\n  const paragraphs = text.split(/\\n\\s*\\n/).filter(p => p.trim());\n\n  let current = '';\n  for (const para of paragraphs) {\n    if (current.length + para.length + 1 <= maxChars) {\n      current = current ? current + '\\n\\n' + para : para;\n    } else {\n      if (current) chunks.push(current);\n      if (para.length > maxChars) {\n        const sentences = para.split(/(?<=[。！？.!?\\n])/).filter(s => s.trim());\n        let sentBuf = '';\n        for (const sent of sentences) {\n          if (sentBuf.length + sent.length <= maxChars) {\n            sentBuf += sent;\n          } else {\n            if (sentBuf) chunks.push(sentBuf);\n            sentBuf = sent;\n          }\n        }\n        if (sentBuf) current = sentBuf;\n        else current = '';\n      } else {\n        current = para;\n      }\n    }\n  }\n  if (current) chunks.push(current);\n\n  if (CHUNK_OVERLAP > 0 && chunks.length > 1) {\n    const overlapped = [chunks[0]];\n    for (let i = 1; i < chunks.length; i++) {\n      const prevTail = chunks[i - 1].slice(-CHUNK_OVERLAP);\n      overlapped.push(prevTail + chunks[i]);\n    }\n    return overlapped;\n  }\n\n  return chunks;\n}\n\n/**\n * 将知识条目拆分成可索引的块\n * 使用字符串复合 ID: \"{source}:{parentId}:{chunkIndex}\" 避免命名空间冲突\n */\nfunction chunkItem(item) {\n  const fullText = item.content || '';\n  const ns = NS[item.source] || 'k';\n\n  if (fullText.length <= CHUNK_MAX_CHARS) {\n    return [{\n      chunkId: `${ns}:${item.id}:0`,\n      source: item.source,\n      parentId: item.id,\n      chunkIndex: 0,\n      title: item.title,\n      content: item.title + ' ' + fullText,\n      category: item.category,\n    }];\n  }\n\n  const chunks = chunkText(fullText);\n  return chunks.map((chunk, i) => ({\n    chunkId: `${ns}:${item.id}:${i}`,\n    source: item.source,\n    parentId: item.id,\n    chunkIndex: i,\n    title: item.title + (chunks.length > 1 ? ` (第${i + 1}/${chunks.length}部分)` : ''),\n    content: item.title + ' ' + chunk,\n    category: item.category,\n  }));\n}\n\n// 块内容缓存（chunkId → {title, content, category, source, parentId}）\nlet chunkContentCache = {};\n\n/**\n * 重建索引（异步）\n * @param {boolean} forceVectors - 强制重建向量\n * @returns {Promise<number>} 索引条目数\n */\nasync function rebuildIndex(forceVectors = false) {\n  const db = getDb();\n\n  // 1. ai_knowledge 表\n  const knowledgeItems = db.prepare(\n    \"SELECT id, title, content, category FROM ai_knowledge WHERE status = 'active'\"\n  ).all().map(item => ({ ...item, source: 'knowledge' }));\n\n  // 2. tutorials 表（已发布教程）\n  const tutorialItems = db.prepare(\n    \"SELECT id, title, content, category FROM tutorials WHERE status = 'published'\"\n  ).all().map(item => ({ ...item, source: 'tutorial' }));\n\n  // 3. faqs 表（激活状态）\n  const faqItems = db.prepare(\n    \"SELECT id, question as title, answer as content, category FROM faqs WHERE status = 'active'\"\n  ).all().map(item => ({ ...item, source: 'faq' }));\n\n  // 合并所有数据源\n  const rawItems = [...knowledgeItems, ...tutorialItems, ...faqItems];\n\n  // 拆分长文档\n  const allItems = [];\n  let chunkCount = 0;\n  for (const item of rawItems) {\n    const chunks = chunkItem(item);\n    allItems.push(...chunks);\n    if (chunks.length > 1) chunkCount++;\n  }\n  if (chunkCount > 0) {\n    console.log(`文档拆分: ${chunkCount} 篇长文档被拆分成 ${allItems.length} 个块`);\n  }\n\n  // 更新块内容缓存\n  chunkContentCache = {};\n  for (const item of allItems) {\n    chunkContentCache[item.chunkId] = {\n      title: item.title,\n      content: item.content,\n      category: item.category,\n      source: item.source,\n      parentId: item.parentId,\n    };\n  }\n\n  // 重建 BM25\n  bm25Index = new BM25Index();\n  bm25Index.build(allItems.map(item => ({ ...item, id: item.chunkId })));\n\n  // 重建向量（使用真实 Embedding API）\n  try {\n    vectorIndex.rebuildTable(db);\n\n    const currentVecCount = db.prepare('SELECT COUNT(*) as cnt FROM knowledge_vec').get().cnt;\n    if (!forceVectors && currentVecCount >= allItems.length) {\n      // 跳过向量重建，但仍需恢复内存映射（重启后 chunkMeta 为空）\n      vectorIdCounter = 0;\n      chunkMeta = {};\n      chunkIdToVectorId = {};\n      for (let i = 0; i < allItems.length; i++) {\n        const vecId = ++vectorIdCounter;\n        chunkMeta[vecId] = {\n          source: allItems[i].source,\n          parentId: allItems[i].parentId,\n          chunkIndex: allItems[i].chunkIndex,\n          chunkId: allItems[i].chunkId,\n        };\n        chunkIdToVectorId[allItems[i].chunkId] = vecId;\n      }\n      console.log(`向量索引已存在 (${currentVecCount} 条)，恢复内存映射`);\n    } else {\n      try { db.prepare('DELETE FROM knowledge_vec').run(); } catch {}\n\n      // 重置映射\n      vectorIdCounter = 0;\n      chunkMeta = {};\n      chunkIdToVectorId = {};\n\n      const texts = allItems.map(item => item.title + ' ' + item.content);\n      const embedConfig = getEmbeddingConfig();\n\n      if (embedConfig.apiKey) {\n        console.log(`正在调用 ${embedConfig.provider} Embedding API，共 ${texts.length} 条...`);\n        const embeddings = await callEmbeddingBatch(texts, embedConfig);\n\n        let successCount = 0;\n        for (let i = 0; i < allItems.length; i++) {\n          const vecId = ++vectorIdCounter;\n          if (embeddings[i] && vectorIndex.upsert(db, vecId, embeddings[i])) {\n            chunkMeta[vecId] = {\n              source: allItems[i].source,\n              parentId: allItems[i].parentId,\n              chunkIndex: allItems[i].chunkIndex,\n              chunkId: allItems[i].chunkId,\n            };\n            chunkIdToVectorId[allItems[i].chunkId] = vecId;\n            successCount++;\n          }\n        }\n        console.log(`向量索引已建立: ${successCount}/${allItems.length} 条成功`);\n      } else {\n        console.warn('未配置 AI API Key，跳过向量索引构建（仅使用 BM25）');\n      }\n    }\n  } catch (e) {\n    console.warn('向量索引初始化失败（不影响 BM25 检索）:', e.message);\n  }\n\n  console.log(`RAG 索引已重建: ${allItems.length} 条 (知识库${knowledgeItems.length} + 教程${tutorialItems.length} + FAQ${faqItems.length})`);\n  return allItems.length;\n}\n\n/**\n * 混合检索：BM25 + 向量，加权合并\n * @param {string} query - 用户问题\n * @param {number} topK - 返回条数\n * @returns {Promise<Array>} [{chunkId, title, content, score, source, parentId}]\n */\nasync function retrieve(query, topK = 5) {\n  const db = getDb();\n\n  if (!bm25Index) await rebuildIndex();\n\n  // BM25 检索\n  const bm25Results = bm25Index.search(query, topK * 2);\n\n  // 向量检索\n  let vecResults = [];\n  try {\n    const embedConfig = getEmbeddingConfig();\n    if (embedConfig.apiKey) {\n      const queryVector = await callEmbeddingAPI(query, embedConfig);\n      if (queryVector) {\n        const rawVecResults = vectorIndex.search(db, queryVector, topK * 2);\n        // 将 vectorId 映射回 chunkId\n        for (const r of rawVecResults) {\n          const meta = chunkMeta[r.vectorId];\n          if (meta) {\n            vecResults.push({ id: meta.chunkId, score: r.score });\n          }\n        }\n      }\n    }\n  } catch {}\n\n  // 合并分数（加权：BM25 0.6 + 向量 0.4）\n  const scoreMap = {};\n  for (const r of bm25Results) {\n    scoreMap[r.id] = (scoreMap[r.id] || 0) + r.score * 0.6;\n  }\n  for (const r of vecResults) {\n    scoreMap[r.id] = (scoreMap[r.id] || 0) + r.score * 0.4;\n  }\n\n  const sorted = Object.entries(scoreMap)\n    .map(([id, score]) => ({ chunkId: id, score }))\n    .sort((a, b) => b.score - a.score)\n    .slice(0, topK);\n\n  if (sorted.length === 0) return [];\n\n  // 获取详情\n  const itemMap = {};\n  const uncachedItems = [];\n\n  for (const s of sorted) {\n    if (chunkContentCache[s.chunkId]) {\n      itemMap[s.chunkId] = chunkContentCache[s.chunkId];\n    } else {\n      uncachedItems.push(s);\n    }\n  }\n\n  // 缓存未命中的，按来源查数据库\n  const knowledgeParents = new Set();\n  const tutorialParents = new Set();\n  const faqParents = new Set();\n\n  for (const s of uncachedItems) {\n    const [source, parentId] = s.chunkId.split(':');\n    const pid = Number(parentId);\n    if (source === 'k') knowledgeParents.add(pid);\n    else if (source === 't') tutorialParents.add(pid);\n    else if (source === 'f') faqParents.add(pid);\n  }\n\n  if (knowledgeParents.size > 0) {\n    const ids = [...knowledgeParents];\n    const ph = ids.map(() => '?').join(',');\n    const rows = db.prepare(`SELECT id, title, content, category FROM ai_knowledge WHERE id IN (${ph}) AND status = 'active'`).all(...ids);\n    for (const r of rows) {\n      itemMap[`k:${r.id}:0`] = { ...r, source: 'knowledge', parentId: r.id };\n    }\n  }\n  if (tutorialParents.size > 0) {\n    const ids = [...tutorialParents];\n    const ph = ids.map(() => '?').join(',');\n    const rows = db.prepare(`SELECT id, title, content, category FROM tutorials WHERE id IN (${ph}) AND status = 'published'`).all(...ids);\n    for (const r of rows) {\n      itemMap[`t:${r.id}:0`] = { ...r, source: 'tutorial', parentId: r.id };\n    }\n  }\n  if (faqParents.size > 0) {\n    const ids = [...faqParents];\n    const ph = ids.map(() => '?').join(',');\n    const rows = db.prepare(`SELECT id, question as title, answer as content, category FROM faqs WHERE id IN (${ph}) AND status = 'active'`).all(...ids);\n    for (const r of rows) {\n      itemMap[`f:${r.id}:0`] = { ...r, source: 'faq', parentId: r.id };\n    }\n  }\n\n  const results = sorted\n    .filter(s => itemMap[s.chunkId])\n    .map(s => {\n      const item = itemMap[s.chunkId];\n      const images = [];\n      const content = item.content || '';\n      const mdImgs = content.match(/!\\[.*?\\]\\((https?:\\/\\/[^)]+)\\)/g) || [];\n      for (const m of mdImgs) {\n        const url = m.match(/\\((https?:\\/\\/[^)]+)\\)/);\n        if (url) images.push(url[1]);\n      }\n      const htmlImgs = content.match(/<img[^>]+src=[\"'](https?:\\/\\/[^\"']+)[\"'][^>]*>/gi) || [];\n      for (const m of htmlImgs) {\n        const url = m.match(/src=[\"'](https?:\\/\\/[^\"']+)[\"']/i);\n        if (url) images.push(url[1]);\n      }\n      return {\n        title: item.title,\n        content: item.content,\n        category: item.category,\n        source: item.source,\n        parentId: item.parentId,\n        score: s.score,\n        images: [...new Set(images)],\n      };\n    });\n\n  // 更新知识库命中次数\n  try {\n    const knowledgeIds = results.filter(r => r.source === 'knowledge').map(r => r.parentId);\n    if (knowledgeIds.length > 0) {\n      const placeholders = knowledgeIds.map(() => '?').join(',');\n      db.prepare(`UPDATE ai_knowledge SET hit_count = hit_count + 1 WHERE id IN (${placeholders})`).run(...knowledgeIds);\n    }\n  } catch {}\n\n  return results;\n}\n\n/**\n * 计算两个文本的相似度\n */\nfunction similarity(textA, textB) {\n  const tokensA = new Set(tokenize(textA));\n  const tokensB = new Set(tokenize(textB));\n  if (tokensA.size === 0 || tokensB.size === 0) return 0;\n  let overlap = 0;\n  for (const t of tokensA) {\n    if (tokensB.has(t)) overlap++;\n  }\n  const union = new Set([...tokensA, ...tokensB]).size;\n  return union > 0 ? overlap / union : 0;\n}\n\n/**\n * 查找与给定问题相似的已有自动学习条目\n */\nasync function findSimilarLearned(question, threshold = 2.0) {\n  const db = getDb();\n\n  if (!bm25Index) await rebuildIndex();\n\n  const bm25Results = bm25Index.search(question, 10);\n\n  for (const result of bm25Results) {\n    if (result.score < threshold) break;\n\n    // chunkId 格式: \"k:{parentId}:{chunkIndex}\"\n    const parts = result.id.split(':');\n    const parentId = Number(parts[1]);\n\n    const item = db.prepare(\n      \"SELECT id, title, content FROM ai_knowledge WHERE id = ? AND category = '自动学习' AND status = 'active'\"\n    ).get(parentId);\n\n    if (item) {\n      return { ...item, score: result.score };\n    }\n  }\n\n  return null;\n}\n\n/**\n * 清理自动学习条目\n */\nfunction cleanupLearned(maxCount = 200) {\n  const db = getDb();\n  const items = db.prepare(\n    \"SELECT id, title, content FROM ai_knowledge WHERE category = '自动学习' AND status = 'active' ORDER BY updated_at DESC\"\n  ).all();\n\n  if (items.length === 0) return { merged: 0, removed: 0 };\n\n  let merged = 0;\n  const toDelete = new Set();\n\n  for (let i = 0; i < items.length; i++) {\n    if (toDelete.has(items[i].id)) continue;\n    for (let j = i + 1; j < items.length; j++) {\n      if (toDelete.has(items[j].id)) continue;\n      const score = similarity(items[i].title, items[j].title);\n      if (score >= 0.4) {\n        const existingContent = items[i].content;\n        const newAnswer = items[j].content.replace(/^用户问题：.*?\\n\\n参考回答：/s, '');\n        if (newAnswer && !existingContent.includes(newAnswer.slice(0, 50))) {\n          db.prepare('UPDATE ai_knowledge SET content = content || ? WHERE id = ?')\n            .run('\\n\\n补充回答：' + newAnswer, items[i].id);\n        }\n        toDelete.add(items[j].id);\n        merged++;\n      }\n    }\n  }\n\n  if (toDelete.size > 0) {\n    const placeholders = Array.from(toDelete).map(() => '?').join(',');\n    db.prepare(`DELETE FROM ai_knowledge WHERE id IN (${placeholders})`).run(...toDelete);\n  }\n\n  let removed = 0;\n  const remaining = db.prepare(\n    \"SELECT COUNT(*) as cnt FROM ai_knowledge WHERE category = '自动学习' AND status = 'active'\"\n  ).get().cnt;\n\n  if (remaining > maxCount) {\n    const excess = remaining - maxCount;\n    const oldItems = db.prepare(\n      \"SELECT id FROM ai_knowledge WHERE category = '自动学习' AND status = 'active' ORDER BY updated_at ASC LIMIT ?\"\n    ).all(excess);\n    if (oldItems.length > 0) {\n      const placeholders = oldItems.map(() => '?').join(',');\n      db.prepare(`DELETE FROM ai_knowledge WHERE id IN (${placeholders})`).run(...oldItems.map(i => i.id));\n      removed = oldItems.length;\n    }\n  }\n\n  return { merged, removed };\n}\n\nmodule.exports = { retrieve, rebuildIndex, tokenize, similarity, findSimilarLearned, cleanupLearned };\n"}, "54c84ea54c3a90233572722cf5e1305947594777": {"t": 1785766725.6997025, "n": 23110, "h": "d9b588fd40ec18f2", "c": "/**\n * RAG 检索服务 — 混合检索（BM25 + 向量）+ Re-ranking\n * 向量检索使用真实 Embedding API（DeepSeek / OpenAI 兼容）\n */\nconst { getDb } = require('../database/schema');\n\nlet sqliteVecLoaded = false;\nfunction ensureVec(db) {\n  if (sqliteVecLoaded) return;\n  try {\n    const sqliteVec = require('sqlite-vec');\n    sqliteVec.load(db);\n    sqliteVecLoaded = true;\n  } catch (e) {\n    console.warn('sqlite-vec 加载失败:', e.message);\n  }\n}\n\n// Embedding 配置\nconst EMBEDDING_DIMENSION = 512; // bge-small-zh-v1.5 输出维度\nconst EMBEDDING_BATCH_SIZE = 10; // DashScope 限制每批最多 10 条\n\n// 命名空间前缀（用于 chunk ID，避免冲突）\nconst NS = { knowledge: 'k', tutorial: 't', faq: 'f' };\n\n// Embedding 模型映射\nconst EMBEDDING_MODELS = {\n  deepseek: 'deepseek-embedding',\n  openai: 'text-embedding-3-small',\n  qwen: 'text-embedding-v3',\n};\n\n// Provider 对应的 Embedding API URL\nfunction getEmbeddingUrl(provider) {\n  const urls = {\n    deepseek: 'https://api.deepseek.com/embeddings',\n    openai: 'https://api.openai.com/v1/embeddings',\n    qwen: 'https://dashscope.aliyuncs.com/compatible-mode/v1/embeddings',\n  };\n  return urls[provider] || '';\n}\n\n// BM25 参数\nconst BM25_K1 = 1.5;\nconst BM25_B = 0.75;\n\n// 同义词映射（用于查询扩展）\nconst SYNONYMS = {\n  '工单': ['工单', '工单', '问题', '反馈', '需求'],\n  '教程': ['教程', '课程', '学习', '视频'],\n  '客服': ['客服', '人工', '帮助', '支持'],\n};\n\n// 中文停用词\nconst STOP_WORDS = new Set([\n  '的', '了', '是', '在', '我', '有', '和', '就', '不', '人', '都', '一',\n  '一个', '上', '也', '很', '到', '说', '要', '去', '你', '会', '着', '没有',\n  '看', '好', '自己', '这', '他', '么', '把', '那', '你', '它', '她', '哪',\n  '如何', '怎么', '什么', '为什么', '可以', '能', '可以', '请', '求', '帮',\n]);\n\n/**\n * 分词（简单实现：中文逐字 + 英文按空格）\n */\nfunction tokenize(text) {\n  if (!text) return [];\n  const tokens = [];\n  // 匹配中文字符或英文单词\n  const matches = text.match(/[\\u4e00-\\u9fff]|[a-zA-Z0-9]+/g) || [];\n  for (const m of matches) {\n    const lower = m.toLowerCase();\n    if (!STOP_WORDS.has(lower) && lower.length > 0) {\n      tokens.push(lower);\n    }\n  }\n  return tokens;\n}\n\n/**\n * 查询扩展：添加同义词\n */\nfunction expandQuery(query) {\n  let expanded = query;\n  for (const [key, synonyms] of Object.entries(SYNONYMS)) {\n    if (query.includes(key)) {\n      expanded += ' ' + synonyms.join(' ');\n    }\n  }\n  return expanded;\n}\n\n// ===== BM25 索引 =====\nclass BM25Index {\n  constructor() {\n    this.docs = [];\n    this.docCount = 0;\n    this.avgDocLen = 0;\n    this.docFreq = {}; // term -> 文档频率\n    this.docLens = [];\n    this.invertedIndex = {}; // term -> [{docIdx, tf}]\n  }\n\n  build(items) {\n    this.docs = items;\n    this.docCount = items.length;\n    this.docFreq = {};\n    this.invertedIndex = {};\n    this.docLens = [];\n\n    let totalLen = 0;\n    for (let i = 0; i < items.length; i++) {\n      const tokens = tokenize(items[i].title + ' ' + items[i].content);\n      this.docLens.push(tokens.length);\n      totalLen += tokens.length;\n\n      // 统计词频\n      const tf = {};\n      for (const t of tokens) tf[t] = (tf[t] || 0) + 1;\n\n      // 构建倒排索引\n      for (const [term, freq] of Object.entries(tf)) {\n        if (!this.invertedIndex[term]) this.invertedIndex[term] = [];\n        this.invertedIndex[term].push({ docIdx: i, tf: freq });\n        this.docFreq[term] = (this.docFreq[term] || 0) + 1;\n      }\n    }\n    this.avgDocLen = totalLen / this.docCount || 1;\n  }\n\n  search(query, topK = 5) {\n    const expandedQuery = expandQuery(query);\n    const queryTokens = tokenize(expandedQuery);\n    if (queryTokens.length === 0) return [];\n\n    const scores = new Array(this.docCount).fill(0);\n\n    for (const term of queryTokens) {\n      const postings = this.invertedIndex[term] || [];\n      const df = this.docFreq[term] || 0;\n      if (df === 0) continue;\n\n      // IDF 部分\n      const idf = Math.log((this.docCount - df + 0.5) / (df + 0.5) + 1);\n\n      for (const { docIdx, tf } of postings) {\n        const docLen = this.docLens[docIdx];\n        // BM25 公式\n        const tfNorm = (tf * (BM25_K1 + 1)) / (tf + BM25_K1 * (1 - BM25_B + BM25_B * docLen / this.avgDocLen));\n        scores[docIdx] += idf * tfNorm;\n      }\n    }\n\n    // 排序取 Top-K\n    const results = [];\n    for (let i = 0; i < this.docCount; i++) {\n      if (scores[i] > 0) {\n        results.push({ id: this.docs[i].id, score: scores[i] });\n      }\n    }\n    results.sort((a, b) => b.score - a.score);\n    return results.slice(0, topK);\n  }\n}\n\n/**\n * 调用 Embedding API（异步，使用 fetch）\n * @param {string} text - 要嵌入的文本\n * @param {object} config - {provider, apiKey, baseUrl, model}\n * @returns {Promise<Float32Array|null>} 嵌入向量\n */\nasync function callEmbeddingAPI(text, config) {\n  const url = config.baseUrl || getEmbeddingUrl(config.provider);\n  if (!url || !config.apiKey) return null;\n\n  const model = config.model || EMBEDDING_MODELS[config.provider] || 'deepseek-embedding';\n\n  try {\n    const resp = await fetch(url, {\n      method: 'POST',\n      headers: {\n        'Content-Type': 'application/json',\n        'Authorization': `Bearer ${config.apiKey}`,\n      },\n      body: JSON.stringify({\n        model,\n        input: text.slice(0, 8000),\n      }),\n      signal: AbortSignal.timeout(30000),\n    });\n\n    if (!resp.ok) {\n      console.warn('Embedding API HTTP 错误:', resp.status);\n      return null;\n    }\n\n    const json = await resp.json();\n    if (json.data && json.data[0] && json.data[0].embedding) {\n      return new Float32Array(json.data[0].embedding);\n    }\n    if (json.error) {\n      console.warn('Embedding API 错误:', json.error.message || json.error);\n    }\n    return null;\n  } catch (e) {\n    console.warn('Embedding API 调用失败:', e.message);\n    return null;\n  }\n}\n\n/**\n * 批量调用 Embedding API（异步）\n * @param {string[]} texts - 文本数组\n * @param {object} config - API 配置\n * @returns {Promise<Float32Array[]>} 嵌入向量数组（失败的用 null 占位）\n */\nasync function callEmbeddingBatch(texts, config) {\n  const url = config.baseUrl || getEmbeddingUrl(config.provider);\n  if (!url || !config.apiKey) return texts.map(() => null);\n\n  const model = config.model || EMBEDDING_MODELS[config.provider] || 'deepseek-embedding';\n  const results = new Array(texts.length).fill(null);\n\n  // 分批处理\n  for (let i = 0; i < texts.length; i += EMBEDDING_BATCH_SIZE) {\n    const batch = texts.slice(i, i + EMBEDDING_BATCH_SIZE);\n\n    try {\n      const resp = await fetch(url, {\n        method: 'POST',\n        headers: {\n          'Content-Type': 'application/json',\n          'Authorization': `Bearer ${config.apiKey}`,\n        },\n        body: JSON.stringify({\n          model,\n          input: batch.map(t => t.slice(0, 8000)),\n        }),\n        signal: AbortSignal.timeout(60000),\n      });\n\n      if (!resp.ok) {\n        console.warn('Embedding 批量 API HTTP 错误:', resp.status);\n        continue;\n      }\n\n      const json = await resp.json();\n      if (json.data) {\n        for (const item of json.data) {\n          results[i + item.index] = new Float32Array(item.embedding);\n        }\n      }\n      if (json.error) {\n        console.warn('Embedding 批量 API 错误:', json.error.message || json.error);\n      }\n    } catch (e) {\n      console.warn('Embedding 批量 API 调用失败:', e.message);\n    }\n\n    // 批次间隔，避免限流\n    if (i + EMBEDDING_BATCH_SIZE < texts.length) {\n      await new Promise(r => setTimeout(r, 1000));\n    }\n  }\n\n  return results;\n}\n\n/**\n * 从 settings 读取 Embedding 配置\n * 支持独立的 embedding 配置，回退到通用 ai 配置\n */\nfunction getEmbeddingConfig() {\n  const db = getDb();\n  const rows = db.prepare(\n    \"SELECT key, value FROM settings WHERE key IN ('embedding_provider','embedding_api_key','embedding_base_url','embedding_model','ai_provider','ai_api_key','ai_base_url')\"\n  ).all();\n  const config = {};\n  for (const row of rows) config[row.key] = row.value;\n  return {\n    provider: config.embedding_provider || config.ai_provider || 'deepseek',\n    apiKey: config.embedding_api_key || config.ai_api_key || '',\n    baseUrl: config.embedding_base_url || '',\n    model: config.embedding_model || '',\n  };\n}\n\n// ===== 向量索引（使用 sqlite-vec）=====\n// 向量表用自增整数 ID，chunkMeta 映射 vectorId → {source, parentId, chunkIndex}\nlet vectorIdCounter = 0;\nlet chunkMeta = {}; // vectorId → {source, parentId, chunkIndex, chunkId}\nlet chunkIdToVectorId = {}; // chunkId → vectorId\n\nclass VectorIndex {\n  constructor() {\n    this.dimension = EMBEDDING_DIMENSION;\n  }\n\n  initTable(db) {\n    ensureVec(db);\n    db.exec(`\n      CREATE VIRTUAL TABLE IF NOT EXISTS knowledge_vec USING vec0(\n        id INTEGER PRIMARY KEY,\n        embedding float[${this.dimension}]\n      );\n    `);\n  }\n\n  rebuildTable(db) {\n    ensureVec(db);\n    try {\n      const info = db.prepare(\"SELECT sql FROM sqlite_master WHERE name = 'knowledge_vec'\").get();\n      if (info && !info.sql.includes(`float[${this.dimension}]`)) {\n        console.log(`向量表维度变化，重建表 (${this.dimension}维)`);\n        db.prepare('DROP TABLE IF EXISTS knowledge_vec').run();\n      }\n    } catch {}\n    this.initTable(db);\n  }\n\n  upsert(db, vectorId, embedding) {\n    if (!embedding) return false;\n    try {\n      db.prepare('DELETE FROM knowledge_vec WHERE id = ?').run(vectorId);\n      db.prepare('INSERT INTO knowledge_vec (id, embedding) VALUES (?, ?)').run(\n        vectorId,\n        Buffer.from(embedding.buffer)\n      );\n      return true;\n    } catch (e) {\n      return false;\n    }\n  }\n\n  search(db, queryVector, topK = 5) {\n    ensureVec(db);\n    if (!queryVector) return [];\n    try {\n      const results = db.prepare(`\n        SELECT id, distance FROM knowledge_vec WHERE embedding MATCH ? ORDER BY distance LIMIT ?\n      `).all(Buffer.from(queryVector.buffer), topK);\n      return results.map(r => ({ vectorId: r.id, score: 1 - r.distance }));\n    } catch (e) {\n      return [];\n    }\n  }\n}\n\n// 全局实例\nlet bm25Index = null;\nconst vectorIndex = new VectorIndex();\n\n// ===== 文档拆分 (Chunking) =====\nconst CHUNK_MAX_CHARS = 400;\nconst CHUNK_OVERLAP = 50;\n\n/**\n * 将长文本按段落拆分成小块\n */\nfunction chunkText(text, maxChars = CHUNK_MAX_CHARS) {\n  if (!text || text.length <= maxChars) return [text];\n\n  const chunks = [];\n  const paragraphs = text.split(/\\n\\s*\\n/).filter(p => p.trim());\n\n  let current = '';\n  for (const para of paragraphs) {\n    if (current.length + para.length + 1 <= maxChars) {\n      current = current ? current + '\\n\\n' + para : para;\n    } else {\n      if (current) chunks.push(current);\n      if (para.length > maxChars) {\n        const sentences = para.split(/(?<=[。！？.!?\\n])/).filter(s => s.trim());\n        let sentBuf = '';\n        for (const sent of sentences) {\n          if (sentBuf.length + sent.length <= maxChars) {\n            sentBuf += sent;\n          } else {\n            if (sentBuf) chunks.push(sentBuf);\n            sentBuf = sent;\n          }\n        }\n        if (sentBuf) current = sentBuf;\n        else current = '';\n      } else {\n        current = para;\n      }\n    }\n  }\n  if (current) chunks.push(current);\n\n  if (CHUNK_OVERLAP > 0 && chunks.length > 1) {\n    const overlapped = [chunks[0]];\n    for (let i = 1; i < chunks.length; i++) {\n      const prevTail = chunks[i - 1].slice(-CHUNK_OVERLAP);\n      overlapped.push(prevTail + chunks[i]);\n    }\n    return overlapped;\n  }\n\n  return chunks;\n}\n\n/**\n * 将知识条目拆分成可索引的块\n * 使用字符串复合 ID: \"{source}:{parentId}:{chunkIndex}\" 避免命名空间冲突\n */\nfunction chunkItem(item) {\n  const fullText = item.content || '';\n  const ns = NS[item.source] || 'k';\n\n  if (fullText.length <= CHUNK_MAX_CHARS) {\n    return [{\n      chunkId: `${ns}:${item.id}:0`,\n      source: item.source,\n      parentId: item.id,\n      chunkIndex: 0,\n      title: item.title,\n      content: item.title + ' ' + fullText,\n      category: item.category,\n    }];\n  }\n\n  const chunks = chunkText(fullText);\n  return chunks.map((chunk, i) => ({\n    chunkId: `${ns}:${item.id}:${i}`,\n    source: item.source,\n    parentId: item.id,\n    chunkIndex: i,\n    title: item.title + (chunks.length > 1 ? ` (第${i + 1}/${chunks.length}部分)` : ''),\n    content: item.title + ' ' + chunk,\n    category: item.category,\n  }));\n}\n\n// 块内容缓存（chunkId → {title, content, category, source, parentId}）\nlet chunkContentCache = {};\n\n/**\n * 重建索引（异步）\n * @param {boolean} forceVectors - 强制重建向量\n * @returns {Promise<number>} 索引条目数\n */\nasync function rebuildIndex(forceVectors = false) {\n  const db = getDb();\n\n  // 1. ai_knowledge 表\n  const knowledgeItems = db.prepare(\n    \"SELECT id, title, content, category FROM ai_knowledge WHERE status = 'active'\"\n  ).all().map(item => ({ ...item, source: 'knowledge' }));\n\n  // 2. tutorials 表（已发布教程）\n  const tutorialItems = db.prepare(\n    \"SELECT id, title, content, category FROM tutorials WHERE status = 'published'\"\n  ).all().map(item => ({ ...item, source: 'tutorial' }));\n\n  // 3. faqs 表（激活状态）\n  const faqItems = db.prepare(\n    \"SELECT id, question as title, answer as content, category FROM faqs WHERE status = 'active'\"\n  ).all().map(item => ({ ...item, source: 'faq' }));\n\n  // 合并所有数据源\n  const rawItems = [...knowledgeItems, ...tutorialItems, ...faqItems];\n\n  // 拆分长文档\n  const allItems = [];\n  let chunkCount = 0;\n  for (const item of rawItems) {\n    const chunks = chunkItem(item);\n    allItems.push(...chunks);\n    if (chunks.length > 1) chunkCount++;\n  }\n  if (chunkCount > 0) {\n    console.log(`文档拆分: ${chunkCount} 篇长文档被拆分成 ${allItems.length} 个块`);\n  }\n\n  // 更新块内容缓存\n  chunkContentCache = {};\n  for (const item of allItems) {\n    chunkContentCache[item.chunkId] = {\n      title: item.title,\n      content: item.content,\n      category: item.category,\n      source: item.source,\n      parentId: item.parentId,\n    };\n  }\n\n  // 重建 BM25\n  bm25Index = new BM25Index();\n  bm25Index.build(allItems.map(item => ({ ...item, id: item.chunkId })));\n\n  // 重建向量（使用真实 Embedding API）\n  try {\n    vectorIndex.rebuildTable(db);\n\n    const currentVecCount = db.prepare('SELECT COUNT(*) as cnt FROM knowledge_vec').get().cnt;\n    if (!forceVectors && currentVecCount >= allItems.length) {\n      // 跳过向量重建，但仍需恢复内存映射（重启后 chunkMeta 为空）\n      vectorIdCounter = 0;\n      chunkMeta = {};\n      chunkIdToVectorId = {};\n      for (let i = 0; i < allItems.length; i++) {\n        const vecId = ++vectorIdCounter;\n        chunkMeta[vecId] = {\n          source: allItems[i].source,\n          parentId: allItems[i].parentId,\n          chunkIndex: allItems[i].chunkIndex,\n          chunkId: allItems[i].chunkId,\n        };\n        chunkIdToVectorId[allItems[i].chunkId] = vecId;\n      }\n      console.log(`向量索引已存在 (${currentVecCount} 条)，恢复内存映射`);\n    } else {\n      try { db.prepare('DELETE FROM knowledge_vec').run(); } catch {}\n\n      // 重置映射\n      vectorIdCounter = 0;\n      chunkMeta = {};\n      chunkIdToVectorId = {};\n\n      const texts = allItems.map(item => item.title + ' ' + item.content);\n      const embedConfig = getEmbeddingConfig();\n\n      if (embedConfig.apiKey) {\n        console.log(`正在调用 ${embedConfig.provider} Embedding API，共 ${texts.length} 条...`);\n        const embeddings = await callEmbeddingBatch(texts, embedConfig);\n\n        let successCount = 0;\n        for (let i = 0; i < allItems.length; i++) {\n          const vecId = ++vectorIdCounter;\n          if (embeddings[i] && vectorIndex.upsert(db, vecId, embeddings[i])) {\n            chunkMeta[vecId] = {\n              source: allItems[i].source,\n              parentId: allItems[i].parentId,\n              chunkIndex: allItems[i].chunkIndex,\n              chunkId: allItems[i].chunkId,\n            };\n            chunkIdToVectorId[allItems[i].chunkId] = vecId;\n            successCount++;\n          }\n        }\n        console.log(`向量索引已建立: ${successCount}/${allItems.length} 条成功`);\n      } else {\n        console.warn('未配置 AI API Key，跳过向量索引构建（仅使用 BM25）');\n      }\n    }\n  } catch (e) {\n    console.warn('向量索引初始化失败（不影响 BM25 检索）:', e.message);\n  }\n\n  console.log(`RAG 索引已重建: ${allItems.length} 条 (知识库${knowledgeItems.length} + 教程${tutorialItems.length} + FAQ${faqItems.length})`);\n  return allItems.length;\n}\n\n/**\n * 混合检索：BM25 + 向量，加权合并\n * @param {string} query - 用户问题\n * @param {number} topK - 返回条数\n * @returns {Promise<Array>} [{chunkId, title, content, score, source, parentId}]\n */\nasync function retrieve(query, topK = 5) {\n  const db = getDb();\n\n  if (!bm25Index) await rebuildIndex();\n\n  // BM25 检索\n  const bm25Results = bm25Index.search(query, topK * 2);\n\n  // 向量检索\n  let vecResults = [];\n  try {\n    const embedConfig = getEmbeddingConfig();\n    if (embedConfig.apiKey) {\n      const queryVector = await callEmbeddingAPI(query, embedConfig);\n      if (queryVector) {\n        const rawVecResults = vectorIndex.search(db, queryVector, topK * 2);\n        // 将 vectorId 映射回 chunkId\n        for (const r of rawVecResults) {\n          const meta = chunkMeta[r.vectorId];\n          if (meta) {\n            vecResults.push({ id: meta.chunkId, score: r.score });\n          }\n        }\n      }\n    }\n  } catch {}\n\n  // 合并分数（加权：BM25 0.6 + 向量 0.4）\n  const scoreMap = {};\n  for (const r of bm25Results) {\n    scoreMap[r.id] = (scoreMap[r.id] || 0) + r.score * 0.6;\n  }\n  for (const r of vecResults) {\n    scoreMap[r.id] = (scoreMap[r.id] || 0) + r.score * 0.4;\n  }\n\n  const sorted = Object.entries(scoreMap)\n    .map(([id, score]) => ({ chunkId: id, score }))\n    .sort((a, b) => b.score - a.score)\n    .slice(0, topK);\n\n  if (sorted.length === 0) return [];\n\n  // 获取详情\n  const itemMap = {};\n  const uncachedItems = [];\n\n  for (const s of sorted) {\n    if (chunkContentCache[s.chunkId]) {\n      itemMap[s.chunkId] = chunkContentCache[s.chunkId];\n    } else {\n      uncachedItems.push(s);\n    }\n  }\n\n  // 缓存未命中的，按来源查数据库\n  const knowledgeParents = new Set();\n  const tutorialParents = new Set();\n  const faqParents = new Set();\n\n  for (const s of uncachedItems) {\n    const [source, parentId] = s.chunkId.split(':');\n    const pid = Number(parentId);\n    if (source === 'k') knowledgeParents.add(pid);\n    else if (source === 't') tutorialParents.add(pid);\n    else if (source === 'f') faqParents.add(pid);\n  }\n\n  if (knowledgeParents.size > 0) {\n    const ids = [...knowledgeParents];\n    const ph = ids.map(() => '?').join(',');\n    const rows = db.prepare(`SELECT id, title, content, category FROM ai_knowledge WHERE id IN (${ph}) AND status = 'active'`).all(...ids);\n    for (const r of rows) {\n      itemMap[`k:${r.id}:0`] = { ...r, source: 'knowledge', parentId: r.id };\n    }\n  }\n  if (tutorialParents.size > 0) {\n    const ids = [...tutorialParents];\n    const ph = ids.map(() => '?').join(',');\n    const rows = db.prepare(`SELECT id, title, content, category FROM tutorials WHERE id IN (${ph}) AND status = 'published'`).all(...ids);\n    for (const r of rows) {\n      itemMap[`t:${r.id}:0`] = { ...r, source: 'tutorial', parentId: r.id };\n    }\n  }\n  if (faqParents.size > 0) {\n    const ids = [...faqParents];\n    const ph = ids.map(() => '?').join(',');\n    const rows = db.prepare(`SELECT id, question as title, answer as content, category FROM faqs WHERE id IN (${ph}) AND status = 'active'`).all(...ids);\n    for (const r of rows) {\n      itemMap[`f:${r.id}:0`] = { ...r, source: 'faq', parentId: r.id };\n    }\n  }\n\n  return sorted\n    .filter(s => itemMap[s.chunkId])\n    .map(s => {\n      const item = itemMap[s.chunkId];\n      const images = [];\n      const content = item.content || '';\n      const mdImgs = content.match(/!\\[.*?\\]\\((https?:\\/\\/[^)]+)\\)/g) || [];\n      for (const m of mdImgs) {\n        const url = m.match(/\\((https?:\\/\\/[^)]+)\\)/);\n        if (url) images.push(url[1]);\n      }\n      const htmlImgs = content.match(/<img[^>]+src=[\"'](https?:\\/\\/[^\"']+)[\"'][^>]*>/gi) || [];\n      for (const m of htmlImgs) {\n        const url = m.match(/src=[\"'](https?:\\/\\/[^\"']+)[\"']/i);\n        if (url) images.push(url[1]);\n      }\n      return {\n        title: item.title,\n        content: item.content,\n        category: item.category,\n        source: item.source,\n        parentId: item.parentId,\n        score: s.score,\n        images: [...new Set(images)],\n      };\n    });\n\n  // 异步更新知识库命中次数（不阻塞返回）\n  try {\n    const knowledgeIds = results.filter(r => r.source === 'knowledge').map(r => r.parentId);\n    if (knowledgeIds.length > 0) {\n      const placeholders = knowledgeIds.map(() => '?').join(',');\n      db.prepare(`UPDATE ai_knowledge SET hit_count = hit_count + 1 WHERE id IN (${placeholders})`).run(...knowledgeIds);\n    }\n  } catch {}\n\n  return results;\n}\n\n/**\n * 计算两个文本的相似度\n */\nfunction similarity(textA, textB) {\n  const tokensA = new Set(tokenize(textA));\n  const tokensB = new Set(tokenize(textB));\n  if (tokensA.size === 0 || tokensB.size === 0) return 0;\n  let overlap = 0;\n  for (const t of tokensA) {\n    if (tokensB.has(t)) overlap++;\n  }\n  const union = new Set([...tokensA, ...tokensB]).size;\n  return union > 0 ? overlap / union : 0;\n}\n\n/**\n * 查找与给定问题相似的已有自动学习条目\n */\nasync function findSimilarLearned(question, threshold = 2.0) {\n  const db = getDb();\n\n  if (!bm25Index) await rebuildIndex();\n\n  const bm25Results = bm25Index.search(question, 10);\n\n  for (const result of bm25Results) {\n    if (result.score < threshold) break;\n\n    // chunkId 格式: \"k:{parentId}:{chunkIndex}\"\n    const parts = result.id.split(':');\n    const parentId = Number(parts[1]);\n\n    const item = db.prepare(\n      \"SELECT id, title, content FROM ai_knowledge WHERE id = ? AND category = '自动学习' AND status = 'active'\"\n    ).get(parentId);\n\n    if (item) {\n      return { ...item, score: result.score };\n    }\n  }\n\n  return null;\n}\n\n/**\n * 清理自动学习条目\n */\nfunction cleanupLearned(maxCount = 200) {\n  const db = getDb();\n  const items = db.prepare(\n    \"SELECT id, title, content FROM ai_knowledge WHERE category = '自动学习' AND status = 'active' ORDER BY updated_at DESC\"\n  ).all();\n\n  if (items.length === 0) return { merged: 0, removed: 0 };\n\n  let merged = 0;\n  const toDelete = new Set();\n\n  for (let i = 0; i < items.length; i++) {\n    if (toDelete.has(items[i].id)) continue;\n    for (let j = i + 1; j < items.length; j++) {\n      if (toDelete.has(items[j].id)) continue;\n      const score = similarity(items[i].title, items[j].title);\n      if (score >= 0.4) {\n        const existingContent = items[i].content;\n        const newAnswer = items[j].content.replace(/^用户问题：.*?\\n\\n参考回答：/s, '');\n        if (newAnswer && !existingContent.includes(newAnswer.slice(0, 50))) {\n          db.prepare('UPDATE ai_knowledge SET content = content || ? WHERE id = ?')\n            .run('\\n\\n补充回答：' + newAnswer, items[i].id);\n        }\n        toDelete.add(items[j].id);\n        merged++;\n      }\n    }\n  }\n\n  if (toDelete.size > 0) {\n    const placeholders = Array.from(toDelete).map(() => '?').join(',');\n    db.prepare(`DELETE FROM ai_knowledge WHERE id IN (${placeholders})`).run(...toDelete);\n  }\n\n  let removed = 0;\n  const remaining = db.prepare(\n    \"SELECT COUNT(*) as cnt FROM ai_knowledge WHERE category = '自动学习' AND status = 'active'\"\n  ).get().cnt;\n\n  if (remaining > maxCount) {\n    const excess = remaining - maxCount;\n    const oldItems = db.prepare(\n      \"SELECT id FROM ai_knowledge WHERE category = '自动学习' AND status = 'active' ORDER BY updated_at ASC LIMIT ?\"\n    ).all(excess);\n    if (oldItems.length > 0) {\n      const placeholders = oldItems.map(() => '?').join(',');\n      db.prepare(`DELETE FROM ai_knowledge WHERE id IN (${placeholders})`).run(...oldItems.map(i => i.id));\n      removed = oldItems.length;\n    }\n  }\n\n  return { merged, removed };\n}\n\nmodule.exports = { retrieve, rebuildIndex, tokenize, similarity, findSimilarLearned, cleanupLearned };\n"}, "74397767c20a14462b73dc91cf277dc8516f92a5": {"t": 1785766727.2267144, "n": 22723, "h": "efc9d29dc0c99ec2", "c": "/**\n * RAG 检索服务 — 混合检索（BM25 + 向量）+ Re-ranking\n * 向量检索使用真实 Embedding API（DeepSeek / OpenAI 兼容）\n */\nconst { getDb } = require('../database/schema');\n\nlet sqliteVecLoaded = false;\nfunction ensureVec(db) {\n  if (sqliteVecLoaded) return;\n  try {\n    const sqliteVec = require('sqlite-vec');\n    sqliteVec.load(db);\n    sqliteVecLoaded = true;\n  } catch (e) {\n    console.warn('sqlite-vec 加载失败:', e.message);\n  }\n}\n\n// Embedding 配置\nconst EMBEDDING_DIMENSION = 512; // bge-small-zh-v1.5 输出维度\nconst EMBEDDING_BATCH_SIZE = 10; // DashScope 限制每批最多 10 条\n\n// 命名空间前缀（用于 chunk ID，避免冲突）\nconst NS = { knowledge: 'k', tutorial: 't', faq: 'f' };\n\n// Embedding 模型映射\nconst EMBEDDING_MODELS = {\n  deepseek: 'deepseek-embedding',\n  openai: 'text-embedding-3-small',\n  qwen: 'text-embedding-v3',\n};\n\n// Provider 对应的 Embedding API URL\nfunction getEmbeddingUrl(provider) {\n  const urls = {\n    deepseek: 'https://api.deepseek.com/embeddings',\n    openai: 'https://api.openai.com/v1/embeddings',\n    qwen: 'https://dashscope.aliyuncs.com/compatible-mode/v1/embeddings',\n  };\n  return urls[provider] || '';\n}\n\n// BM25 参数\nconst BM25_K1 = 1.5;\nconst BM25_B = 0.75;\n\n// 同义词映射（用于查询扩展）\nconst SYNONYMS = {\n  '工单': ['工单', '工单', '问题', '反馈', '需求'],\n  '教程': ['教程', '课程', '学习', '视频'],\n  '客服': ['客服', '人工', '帮助', '支持'],\n};\n\n// 中文停用词\nconst STOP_WORDS = new Set([\n  '的', '了', '是', '在', '我', '有', '和', '就', '不', '人', '都', '一',\n  '一个', '上', '也', '很', '到', '说', '要', '去', '你', '会', '着', '没有',\n  '看', '好', '自己', '这', '他', '么', '把', '那', '你', '它', '她', '哪',\n  '如何', '怎么', '什么', '为什么', '可以', '能', '可以', '请', '求', '帮',\n]);\n\n/**\n * 分词（简单实现：中文逐字 + 英文按空格）\n */\nfunction tokenize(text) {\n  if (!text) return [];\n  const tokens = [];\n  // 匹配中文字符或英文单词\n  const matches = text.match(/[\\u4e00-\\u9fff]|[a-zA-Z0-9]+/g) || [];\n  for (const m of matches) {\n    const lower = m.toLowerCase();\n    if (!STOP_WORDS.has(lower) && lower.length > 0) {\n      tokens.push(lower);\n    }\n  }\n  return tokens;\n}\n\n/**\n * 查询扩展：添加同义词\n */\nfunction expandQuery(query) {\n  let expanded = query;\n  for (const [key, synonyms] of Object.entries(SYNONYMS)) {\n    if (query.includes(key)) {\n      expanded += ' ' + synonyms.join(' ');\n    }\n  }\n  return expanded;\n}\n\n// ===== BM25 索引 =====\nclass BM25Index {\n  constructor() {\n    this.docs = [];\n    this.docCount = 0;\n    this.avgDocLen = 0;\n    this.docFreq = {}; // term -> 文档频率\n    this.docLens = [];\n    this.invertedIndex = {}; // term -> [{docIdx, tf}]\n  }\n\n  build(items) {\n    this.docs = items;\n    this.docCount = items.length;\n    this.docFreq = {};\n    this.invertedIndex = {};\n    this.docLens = [];\n\n    let totalLen = 0;\n    for (let i = 0; i < items.length; i++) {\n      const tokens = tokenize(items[i].title + ' ' + items[i].content);\n      this.docLens.push(tokens.length);\n      totalLen += tokens.length;\n\n      // 统计词频\n      const tf = {};\n      for (const t of tokens) tf[t] = (tf[t] || 0) + 1;\n\n      // 构建倒排索引\n      for (const [term, freq] of Object.entries(tf)) {\n        if (!this.invertedIndex[term]) this.invertedIndex[term] = [];\n        this.invertedIndex[term].push({ docIdx: i, tf: freq });\n        this.docFreq[term] = (this.docFreq[term] || 0) + 1;\n      }\n    }\n    this.avgDocLen = totalLen / this.docCount || 1;\n  }\n\n  search(query, topK = 5) {\n    const expandedQuery = expandQuery(query);\n    const queryTokens = tokenize(expandedQuery);\n    if (queryTokens.length === 0) return [];\n\n    const scores = new Array(this.docCount).fill(0);\n\n    for (const term of queryTokens) {\n      const postings = this.invertedIndex[term] || [];\n      const df = this.docFreq[term] || 0;\n      if (df === 0) continue;\n\n      // IDF 部分\n      const idf = Math.log((this.docCount - df + 0.5) / (df + 0.5) + 1);\n\n      for (const { docIdx, tf } of postings) {\n        const docLen = this.docLens[docIdx];\n        // BM25 公式\n        const tfNorm = (tf * (BM25_K1 + 1)) / (tf + BM25_K1 * (1 - BM25_B + BM25_B * docLen / this.avgDocLen));\n        scores[docIdx] += idf * tfNorm;\n      }\n    }\n\n    // 排序取 Top-K\n    const results = [];\n    for (let i = 0; i < this.docCount; i++) {\n      if (scores[i] > 0) {\n        results.push({ id: this.docs[i].id, score: scores[i] });\n      }\n    }\n    results.sort((a, b) => b.score - a.score);\n    return results.slice(0, topK);\n  }\n}\n\n/**\n * 调用 Embedding API（异步，使用 fetch）\n * @param {string} text - 要嵌入的文本\n * @param {object} config - {provider, apiKey, baseUrl, model}\n * @returns {Promise<Float32Array|null>} 嵌入向量\n */\nasync function callEmbeddingAPI(text, config) {\n  const url = config.baseUrl || getEmbeddingUrl(config.provider);\n  if (!url || !config.apiKey) return null;\n\n  const model = config.model || EMBEDDING_MODELS[config.provider] || 'deepseek-embedding';\n\n  try {\n    const resp = await fetch(url, {\n      method: 'POST',\n      headers: {\n        'Content-Type': 'application/json',\n        'Authorization': `Bearer ${config.apiKey}`,\n      },\n      body: JSON.stringify({\n        model,\n        input: text.slice(0, 8000),\n      }),\n      signal: AbortSignal.timeout(30000),\n    });\n\n    if (!resp.ok) {\n      console.warn('Embedding API HTTP 错误:', resp.status);\n      return null;\n    }\n\n    const json = await resp.json();\n    if (json.data && json.data[0] && json.data[0].embedding) {\n      return new Float32Array(json.data[0].embedding);\n    }\n    if (json.error) {\n      console.warn('Embedding API 错误:', json.error.message || json.error);\n    }\n    return null;\n  } catch (e) {\n    console.warn('Embedding API 调用失败:', e.message);\n    return null;\n  }\n}\n\n/**\n * 批量调用 Embedding API（异步）\n * @param {string[]} texts - 文本数组\n * @param {object} config - API 配置\n * @returns {Promise<Float32Array[]>} 嵌入向量数组（失败的用 null 占位）\n */\nasync function callEmbeddingBatch(texts, config) {\n  const url = config.baseUrl || getEmbeddingUrl(config.provider);\n  if (!url || !config.apiKey) return texts.map(() => null);\n\n  const model = config.model || EMBEDDING_MODELS[config.provider] || 'deepseek-embedding';\n  const results = new Array(texts.length).fill(null);\n\n  // 分批处理\n  for (let i = 0; i < texts.length; i += EMBEDDING_BATCH_SIZE) {\n    const batch = texts.slice(i, i + EMBEDDING_BATCH_SIZE);\n\n    try {\n      const resp = await fetch(url, {\n        method: 'POST',\n        headers: {\n          'Content-Type': 'application/json',\n          'Authorization': `Bearer ${config.apiKey}`,\n        },\n        body: JSON.stringify({\n          model,\n          input: batch.map(t => t.slice(0, 8000)),\n        }),\n        signal: AbortSignal.timeout(60000),\n      });\n\n      if (!resp.ok) {\n        console.warn('Embedding 批量 API HTTP 错误:', resp.status);\n        continue;\n      }\n\n      const json = await resp.json();\n      if (json.data) {\n        for (const item of json.data) {\n          results[i + item.index] = new Float32Array(item.embedding);\n        }\n      }\n      if (json.error) {\n        console.warn('Embedding 批量 API 错误:', json.error.message || json.error);\n      }\n    } catch (e) {\n      console.warn('Embedding 批量 API 调用失败:', e.message);\n    }\n\n    // 批次间隔，避免限流\n    if (i + EMBEDDING_BATCH_SIZE < texts.length) {\n      await new Promise(r => setTimeout(r, 1000));\n    }\n  }\n\n  return results;\n}\n\n/**\n * 从 settings 读取 Embedding 配置\n * 支持独立的 embedding 配置，回退到通用 ai 配置\n */\nfunction getEmbeddingConfig() {\n  const db = getDb();\n  const rows = db.prepare(\n    \"SELECT key, value FROM settings WHERE key IN ('embedding_provider','embedding_api_key','embedding_base_url','embedding_model','ai_provider','ai_api_key','ai_base_url')\"\n  ).all();\n  const config = {};\n  for (const row of rows) config[row.key] = row.value;\n  return {\n    provider: config.embedding_provider || config.ai_provider || 'deepseek',\n    apiKey: config.embedding_api_key || config.ai_api_key || '',\n    baseUrl: config.embedding_base_url || '',\n    model: config.embedding_model || '',\n  };\n}\n\n// ===== 向量索引（使用 sqlite-vec）=====\n// 向量表用自增整数 ID，chunkMeta 映射 vectorId → {source, parentId, chunkIndex}\nlet vectorIdCounter = 0;\nlet chunkMeta = {}; // vectorId → {source, parentId, chunkIndex, chunkId}\nlet chunkIdToVectorId = {}; // chunkId → vectorId\n\nclass VectorIndex {\n  constructor() {\n    this.dimension = EMBEDDING_DIMENSION;\n  }\n\n  initTable(db) {\n    ensureVec(db);\n    db.exec(`\n      CREATE VIRTUAL TABLE IF NOT EXISTS knowledge_vec USING vec0(\n        id INTEGER PRIMARY KEY,\n        embedding float[${this.dimension}]\n      );\n    `);\n  }\n\n  rebuildTable(db) {\n    ensureVec(db);\n    try {\n      const info = db.prepare(\"SELECT sql FROM sqlite_master WHERE name = 'knowledge_vec'\").get();\n      if (info && !info.sql.includes(`float[${this.dimension}]`)) {\n        console.log(`向量表维度变化，重建表 (${this.dimension}维)`);\n        db.prepare('DROP TABLE IF EXISTS knowledge_vec').run();\n      }\n    } catch {}\n    this.initTable(db);\n  }\n\n  upsert(db, vectorId, embedding) {\n    if (!embedding) return false;\n    try {\n      db.prepare('DELETE FROM knowledge_vec WHERE id = ?').run(vectorId);\n      db.prepare('INSERT INTO knowledge_vec (id, embedding) VALUES (?, ?)').run(\n        vectorId,\n        Buffer.from(embedding.buffer)\n      );\n      return true;\n    } catch (e) {\n      return false;\n    }\n  }\n\n  search(db, queryVector, topK = 5) {\n    ensureVec(db);\n    if (!queryVector) return [];\n    try {\n      const results = db.prepare(`\n        SELECT id, distance FROM knowledge_vec WHERE embedding MATCH ? ORDER BY distance LIMIT ?\n      `).all(Buffer.from(queryVector.buffer), topK);\n      return results.map(r => ({ vectorId: r.id, score: 1 - r.distance }));\n    } catch (e) {\n      return [];\n    }\n  }\n}\n\n// 全局实例\nlet bm25Index = null;\nconst vectorIndex = new VectorIndex();\n\n// ===== 文档拆分 (Chunking) =====\nconst CHUNK_MAX_CHARS = 400;\nconst CHUNK_OVERLAP = 50;\n\n/**\n * 将长文本按段落拆分成小块\n */\nfunction chunkText(text, maxChars = CHUNK_MAX_CHARS) {\n  if (!text || text.length <= maxChars) return [text];\n\n  const chunks = [];\n  const paragraphs = text.split(/\\n\\s*\\n/).filter(p => p.trim());\n\n  let current = '';\n  for (const para of paragraphs) {\n    if (current.length + para.length + 1 <= maxChars) {\n      current = current ? current + '\\n\\n' + para : para;\n    } else {\n      if (current) chunks.push(current);\n      if (para.length > maxChars) {\n        const sentences = para.split(/(?<=[。！？.!?\\n])/).filter(s => s.trim());\n        let sentBuf = '';\n        for (const sent of sentences) {\n          if (sentBuf.length + sent.length <= maxChars) {\n            sentBuf += sent;\n          } else {\n            if (sentBuf) chunks.push(sentBuf);\n            sentBuf = sent;\n          }\n        }\n        if (sentBuf) current = sentBuf;\n        else current = '';\n      } else {\n        current = para;\n      }\n    }\n  }\n  if (current) chunks.push(current);\n\n  if (CHUNK_OVERLAP > 0 && chunks.length > 1) {\n    const overlapped = [chunks[0]];\n    for (let i = 1; i < chunks.length; i++) {\n      const prevTail = chunks[i - 1].slice(-CHUNK_OVERLAP);\n      overlapped.push(prevTail + chunks[i]);\n    }\n    return overlapped;\n  }\n\n  return chunks;\n}\n\n/**\n * 将知识条目拆分成可索引的块\n * 使用字符串复合 ID: \"{source}:{parentId}:{chunkIndex}\" 避免命名空间冲突\n */\nfunction chunkItem(item) {\n  const fullText = item.content || '';\n  const ns = NS[item.source] || 'k';\n\n  if (fullText.length <= CHUNK_MAX_CHARS) {\n    return [{\n      chunkId: `${ns}:${item.id}:0`,\n      source: item.source,\n      parentId: item.id,\n      chunkIndex: 0,\n      title: item.title,\n      content: item.title + ' ' + fullText,\n      category: item.category,\n    }];\n  }\n\n  const chunks = chunkText(fullText);\n  return chunks.map((chunk, i) => ({\n    chunkId: `${ns}:${item.id}:${i}`,\n    source: item.source,\n    parentId: item.id,\n    chunkIndex: i,\n    title: item.title + (chunks.length > 1 ? ` (第${i + 1}/${chunks.length}部分)` : ''),\n    content: item.title + ' ' + chunk,\n    category: item.category,\n  }));\n}\n\n// 块内容缓存（chunkId → {title, content, category, source, parentId}）\nlet chunkContentCache = {};\n\n/**\n * 重建索引（异步）\n * @param {boolean} forceVectors - 强制重建向量\n * @returns {Promise<number>} 索引条目数\n */\nasync function rebuildIndex(forceVectors = false) {\n  const db = getDb();\n\n  // 1. ai_knowledge 表\n  const knowledgeItems = db.prepare(\n    \"SELECT id, title, content, category FROM ai_knowledge WHERE status = 'active'\"\n  ).all().map(item => ({ ...item, source: 'knowledge' }));\n\n  // 2. tutorials 表（已发布教程）\n  const tutorialItems = db.prepare(\n    \"SELECT id, title, content, category FROM tutorials WHERE status = 'published'\"\n  ).all().map(item => ({ ...item, source: 'tutorial' }));\n\n  // 3. faqs 表（激活状态）\n  const faqItems = db.prepare(\n    \"SELECT id, question as title, answer as content, category FROM faqs WHERE status = 'active'\"\n  ).all().map(item => ({ ...item, source: 'faq' }));\n\n  // 合并所有数据源\n  const rawItems = [...knowledgeItems, ...tutorialItems, ...faqItems];\n\n  // 拆分长文档\n  const allItems = [];\n  let chunkCount = 0;\n  for (const item of rawItems) {\n    const chunks = chunkItem(item);\n    allItems.push(...chunks);\n    if (chunks.length > 1) chunkCount++;\n  }\n  if (chunkCount > 0) {\n    console.log(`文档拆分: ${chunkCount} 篇长文档被拆分成 ${allItems.length} 个块`);\n  }\n\n  // 更新块内容缓存\n  chunkContentCache = {};\n  for (const item of allItems) {\n    chunkContentCache[item.chunkId] = {\n      title: item.title,\n      content: item.content,\n      category: item.category,\n      source: item.source,\n      parentId: item.parentId,\n    };\n  }\n\n  // 重建 BM25\n  bm25Index = new BM25Index();\n  bm25Index.build(allItems.map(item => ({ ...item, id: item.chunkId })));\n\n  // 重建向量（使用真实 Embedding API）\n  try {\n    vectorIndex.rebuildTable(db);\n\n    const currentVecCount = db.prepare('SELECT COUNT(*) as cnt FROM knowledge_vec').get().cnt;\n    if (!forceVectors && currentVecCount >= allItems.length) {\n      // 跳过向量重建，但仍需恢复内存映射（重启后 chunkMeta 为空）\n      vectorIdCounter = 0;\n      chunkMeta = {};\n      chunkIdToVectorId = {};\n      for (let i = 0; i < allItems.length; i++) {\n        const vecId = ++vectorIdCounter;\n        chunkMeta[vecId] = {\n          source: allItems[i].source,\n          parentId: allItems[i].parentId,\n          chunkIndex: allItems[i].chunkIndex,\n          chunkId: allItems[i].chunkId,\n        };\n        chunkIdToVectorId[allItems[i].chunkId] = vecId;\n      }\n      console.log(`向量索引已存在 (${currentVecCount} 条)，恢复内存映射`);\n    } else {\n      try { db.prepare('DELETE FROM knowledge_vec').run(); } catch {}\n\n      // 重置映射\n      vectorIdCounter = 0;\n      chunkMeta = {};\n      chunkIdToVectorId = {};\n\n      const texts = allItems.map(item => item.title + ' ' + item.content);\n      const embedConfig = getEmbeddingConfig();\n\n      if (embedConfig.apiKey) {\n        console.log(`正在调用 ${embedConfig.provider} Embedding API，共 ${texts.length} 条...`);\n        const embeddings = await callEmbeddingBatch(texts, embedConfig);\n\n        let successCount = 0;\n        for (let i = 0; i < allItems.length; i++) {\n          const vecId = ++vectorIdCounter;\n          if (embeddings[i] && vectorIndex.upsert(db, vecId, embeddings[i])) {\n            chunkMeta[vecId] = {\n              source: allItems[i].source,\n              parentId: allItems[i].parentId,\n              chunkIndex: allItems[i].chunkIndex,\n              chunkId: allItems[i].chunkId,\n            };\n            chunkIdToVectorId[allItems[i].chunkId] = vecId;\n            successCount++;\n          }\n        }\n        console.log(`向量索引已建立: ${successCount}/${allItems.length} 条成功`);\n      } else {\n        console.warn('未配置 AI API Key，跳过向量索引构建（仅使用 BM25）');\n      }\n    }\n  } catch (e) {\n    console.warn('向量索引初始化失败（不影响 BM25 检索）:', e.message);\n  }\n\n  console.log(`RAG 索引已重建: ${allItems.length} 条 (知识库${knowledgeItems.length} + 教程${tutorialItems.length} + FAQ${faqItems.length})`);\n  return allItems.length;\n}\n\n/**\n * 混合检索：BM25 + 向量，加权合并\n * @param {string} query - 用户问题\n * @param {number} topK - 返回条数\n * @returns {Promise<Array>} [{chunkId, title, content, score, source, parentId}]\n */\nasync function retrieve(query, topK = 5) {\n  const db = getDb();\n\n  if (!bm25Index) await rebuildIndex();\n\n  // BM25 检索\n  const bm25Results = bm25Index.search(query, topK * 2);\n\n  // 向量检索\n  let vecResults = [];\n  try {\n    const embedConfig = getEmbeddingConfig();\n    if (embedConfig.apiKey) {\n      const queryVector = await callEmbeddingAPI(query, embedConfig);\n      if (queryVector) {\n        const rawVecResults = vectorIndex.search(db, queryVector, topK * 2);\n        // 将 vectorId 映射回 chunkId\n        for (const r of rawVecResults) {\n          const meta = chunkMeta[r.vectorId];\n          if (meta) {\n            vecResults.push({ id: meta.chunkId, score: r.score });\n          }\n        }\n      }\n    }\n  } catch {}\n\n  // 合并分数（加权：BM25 0.6 + 向量 0.4）\n  const scoreMap = {};\n  for (const r of bm25Results) {\n    scoreMap[r.id] = (scoreMap[r.id] || 0) + r.score * 0.6;\n  }\n  for (const r of vecResults) {\n    scoreMap[r.id] = (scoreMap[r.id] || 0) + r.score * 0.4;\n  }\n\n  const sorted = Object.entries(scoreMap)\n    .map(([id, score]) => ({ chunkId: id, score }))\n    .sort((a, b) => b.score - a.score)\n    .slice(0, topK);\n\n  if (sorted.length === 0) return [];\n\n  // 获取详情\n  const itemMap = {};\n  const uncachedItems = [];\n\n  for (const s of sorted) {\n    if (chunkContentCache[s.chunkId]) {\n      itemMap[s.chunkId] = chunkContentCache[s.chunkId];\n    } else {\n      uncachedItems.push(s);\n    }\n  }\n\n  // 缓存未命中的，按来源查数据库\n  const knowledgeParents = new Set();\n  const tutorialParents = new Set();\n  const faqParents = new Set();\n\n  for (const s of uncachedItems) {\n    const [source, parentId] = s.chunkId.split(':');\n    const pid = Number(parentId);\n    if (source === 'k') knowledgeParents.add(pid);\n    else if (source === 't') tutorialParents.add(pid);\n    else if (source === 'f') faqParents.add(pid);\n  }\n\n  if (knowledgeParents.size > 0) {\n    const ids = [...knowledgeParents];\n    const ph = ids.map(() => '?').join(',');\n    const rows = db.prepare(`SELECT id, title, content, category FROM ai_knowledge WHERE id IN (${ph}) AND status = 'active'`).all(...ids);\n    for (const r of rows) {\n      itemMap[`k:${r.id}:0`] = { ...r, source: 'knowledge', parentId: r.id };\n    }\n  }\n  if (tutorialParents.size > 0) {\n    const ids = [...tutorialParents];\n    const ph = ids.map(() => '?').join(',');\n    const rows = db.prepare(`SELECT id, title, content, category FROM tutorials WHERE id IN (${ph}) AND status = 'published'`).all(...ids);\n    for (const r of rows) {\n      itemMap[`t:${r.id}:0`] = { ...r, source: 'tutorial', parentId: r.id };\n    }\n  }\n  if (faqParents.size > 0) {\n    const ids = [...faqParents];\n    const ph = ids.map(() => '?').join(',');\n    const rows = db.prepare(`SELECT id, question as title, answer as content, category FROM faqs WHERE id IN (${ph}) AND status = 'active'`).all(...ids);\n    for (const r of rows) {\n      itemMap[`f:${r.id}:0`] = { ...r, source: 'faq', parentId: r.id };\n    }\n  }\n\n  return sorted\n    .filter(s => itemMap[s.chunkId])\n    .map(s => {\n      const item = itemMap[s.chunkId];\n      const images = [];\n      const content = item.content || '';\n      const mdImgs = content.match(/!\\[.*?\\]\\((https?:\\/\\/[^)]+)\\)/g) || [];\n      for (const m of mdImgs) {\n        const url = m.match(/\\((https?:\\/\\/[^)]+)\\)/);\n        if (url) images.push(url[1]);\n      }\n      const htmlImgs = content.match(/<img[^>]+src=[\"'](https?:\\/\\/[^\"']+)[\"'][^>]*>/gi) || [];\n      for (const m of htmlImgs) {\n        const url = m.match(/src=[\"'](https?:\\/\\/[^\"']+)[\"']/i);\n        if (url) images.push(url[1]);\n      }\n      return {\n        title: item.title,\n        content: item.content,\n        category: item.category,\n        source: item.source,\n        parentId: item.parentId,\n        score: s.score,\n        images: [...new Set(images)],\n      };\n    });\n}\n\n/**\n * 计算两个文本的相似度\n */\nfunction similarity(textA, textB) {\n  const tokensA = new Set(tokenize(textA));\n  const tokensB = new Set(tokenize(textB));\n  if (tokensA.size === 0 || tokensB.size === 0) return 0;\n  let overlap = 0;\n  for (const t of tokensA) {\n    if (tokensB.has(t)) overlap++;\n  }\n  const union = new Set([...tokensA, ...tokensB]).size;\n  return union > 0 ? overlap / union : 0;\n}\n\n/**\n * 查找与给定问题相似的已有自动学习条目\n */\nasync function findSimilarLearned(question, threshold = 2.0) {\n  const db = getDb();\n\n  if (!bm25Index) await rebuildIndex();\n\n  const bm25Results = bm25Index.search(question, 10);\n\n  for (const result of bm25Results) {\n    if (result.score < threshold) break;\n\n    // chunkId 格式: \"k:{parentId}:{chunkIndex}\"\n    const parts = result.id.split(':');\n    const parentId = Number(parts[1]);\n\n    const item = db.prepare(\n      \"SELECT id, title, content FROM ai_knowledge WHERE id = ? AND category = '自动学习' AND status = 'active'\"\n    ).get(parentId);\n\n    if (item) {\n      return { ...item, score: result.score };\n    }\n  }\n\n  return null;\n}\n\n/**\n * 清理自动学习条目\n */\nfunction cleanupLearned(maxCount = 200) {\n  const db = getDb();\n  const items = db.prepare(\n    \"SELECT id, title, content FROM ai_knowledge WHERE category = '自动学习' AND status = 'active' ORDER BY updated_at DESC\"\n  ).all();\n\n  if (items.length === 0) return { merged: 0, removed: 0 };\n\n  let merged = 0;\n  const toDelete = new Set();\n\n  for (let i = 0; i < items.length; i++) {\n    if (toDelete.has(items[i].id)) continue;\n    for (let j = i + 1; j < items.length; j++) {\n      if (toDelete.has(items[j].id)) continue;\n      const score = similarity(items[i].title, items[j].title);\n      if (score >= 0.4) {\n        const existingContent = items[i].content;\n        const newAnswer = items[j].content.replace(/^用户问题：.*?\\n\\n参考回答：/s, '');\n        if (newAnswer && !existingContent.includes(newAnswer.slice(0, 50))) {\n          db.prepare('UPDATE ai_knowledge SET content = content || ? WHERE id = ?')\n            .run('\\n\\n补充回答：' + newAnswer, items[i].id);\n        }\n        toDelete.add(items[j].id);\n        merged++;\n      }\n    }\n  }\n\n  if (toDelete.size > 0) {\n    const placeholders = Array.from(toDelete).map(() => '?').join(',');\n    db.prepare(`DELETE FROM ai_knowledge WHERE id IN (${placeholders})`).run(...toDelete);\n  }\n\n  let removed = 0;\n  const remaining = db.prepare(\n    \"SELECT COUNT(*) as cnt FROM ai_knowledge WHERE category = '自动学习' AND status = 'active'\"\n  ).get().cnt;\n\n  if (remaining > maxCount) {\n    const excess = remaining - maxCount;\n    const oldItems = db.prepare(\n      \"SELECT id FROM ai_knowledge WHERE category = '自动学习' AND status = 'active' ORDER BY updated_at ASC LIMIT ?\"\n    ).all(excess);\n    if (oldItems.length > 0) {\n      const placeholders = oldItems.map(() => '?').join(',');\n      db.prepare(`DELETE FROM ai_knowledge WHERE id IN (${placeholders})`).run(...oldItems.map(i => i.id));\n      removed = oldItems.length;\n    }\n  }\n\n  return { merged, removed };\n}\n\nmodule.exports = { retrieve, rebuildIndex, tokenize, similarity, findSimilarLearned, cleanupLearned };\n"}, "07d9a4cb735c78690d56faa55797451246cb6559": {"t": 1785766728.5190501, "n": 22237, "h": "e1f7cc0223611228", "c": "/**\n * RAG 检索服务 — 混合检索（BM25 + 向量）+ Re-ranking\n * 向量检索使用真实 Embedding API（DeepSeek / OpenAI 兼容）\n */\nconst { getDb } = require('../database/schema');\n\nlet sqliteVecLoaded = false;\nfunction ensureVec(db) {\n  if (sqliteVecLoaded) return;\n  try {\n    const sqliteVec = require('sqlite-vec');\n    sqliteVec.load(db);\n    sqliteVecLoaded = true;\n  } catch (e) {\n    console.warn('sqlite-vec 加载失败:', e.message);\n  }\n}\n\n// Embedding 配置\nconst EMBEDDING_DIMENSION = 512; // bge-small-zh-v1.5 输出维度\nconst EMBEDDING_BATCH_SIZE = 10; // DashScope 限制每批最多 10 条\n\n// 命名空间前缀（用于 chunk ID，避免冲突）\nconst NS = { knowledge: 'k', tutorial: 't', faq: 'f' };\n\n// Embedding 模型映射\nconst EMBEDDING_MODELS = {\n  deepseek: 'deepseek-embedding',\n  openai: 'text-embedding-3-small',\n  qwen: 'text-embedding-v3',\n};\n\n// Provider 对应的 Embedding API URL\nfunction getEmbeddingUrl(provider) {\n  const urls = {\n    deepseek: 'https://api.deepseek.com/embeddings',\n    openai: 'https://api.openai.com/v1/embeddings',\n    qwen: 'https://dashscope.aliyuncs.com/compatible-mode/v1/embeddings',\n  };\n  return urls[provider] || '';\n}\n\n// BM25 参数\nconst BM25_K1 = 1.5;\nconst BM25_B = 0.75;\n\n// 同义词映射（用于查询扩展）\nconst SYNONYMS = {\n  '工单': ['工单', '工单', '问题', '反馈', '需求'],\n  '教程': ['教程', '课程', '学习', '视频'],\n  '客服': ['客服', '人工', '帮助', '支持'],\n};\n\n// 中文停用词\nconst STOP_WORDS = new Set([\n  '的', '了', '是', '在', '我', '有', '和', '就', '不', '人', '都', '一',\n  '一个', '上', '也', '很', '到', '说', '要', '去', '你', '会', '着', '没有',\n  '看', '好', '自己', '这', '他', '么', '把', '那', '你', '它', '她', '哪',\n  '如何', '怎么', '什么', '为什么', '可以', '能', '可以', '请', '求', '帮',\n]);\n\n/**\n * 分词（简单实现：中文逐字 + 英文按空格）\n */\nfunction tokenize(text) {\n  if (!text) return [];\n  const tokens = [];\n  // 匹配中文字符或英文单词\n  const matches = text.match(/[\\u4e00-\\u9fff]|[a-zA-Z0-9]+/g) || [];\n  for (const m of matches) {\n    const lower = m.toLowerCase();\n    if (!STOP_WORDS.has(lower) && lower.length > 0) {\n      tokens.push(lower);\n    }\n  }\n  return tokens;\n}\n\n/**\n * 查询扩展：添加同义词\n */\nfunction expandQuery(query) {\n  let expanded = query;\n  for (const [key, synonyms] of Object.entries(SYNONYMS)) {\n    if (query.includes(key)) {\n      expanded += ' ' + synonyms.join(' ');\n    }\n  }\n  return expanded;\n}\n\n// ===== BM25 索引 =====\nclass BM25Index {\n  constructor() {\n    this.docs = [];\n    this.docCount = 0;\n    this.avgDocLen = 0;\n    this.docFreq = {}; // term -> 文档频率\n    this.docLens = [];\n    this.invertedIndex = {}; // term -> [{docIdx, tf}]\n  }\n\n  build(items) {\n    this.docs = items;\n    this.docCount = items.length;\n    this.docFreq = {};\n    this.invertedIndex = {};\n    this.docLens = [];\n\n    let totalLen = 0;\n    for (let i = 0; i < items.length; i++) {\n      const tokens = tokenize(items[i].title + ' ' + items[i].content);\n      this.docLens.push(tokens.length);\n      totalLen += tokens.length;\n\n      // 统计词频\n      const tf = {};\n      for (const t of tokens) tf[t] = (tf[t] || 0) + 1;\n\n      // 构建倒排索引\n      for (const [term, freq] of Object.entries(tf)) {\n        if (!this.invertedIndex[term]) this.invertedIndex[term] = [];\n        this.invertedIndex[term].push({ docIdx: i, tf: freq });\n        this.docFreq[term] = (this.docFreq[term] || 0) + 1;\n      }\n    }\n    this.avgDocLen = totalLen / this.docCount || 1;\n  }\n\n  search(query, topK = 5) {\n    const expandedQuery = expandQuery(query);\n    const queryTokens = tokenize(expandedQuery);\n    if (queryTokens.length === 0) return [];\n\n    const scores = new Array(this.docCount).fill(0);\n\n    for (const term of queryTokens) {\n      const postings = this.invertedIndex[term] || [];\n      const df = this.docFreq[term] || 0;\n      if (df === 0) continue;\n\n      // IDF 部分\n      const idf = Math.log((this.docCount - df + 0.5) / (df + 0.5) + 1);\n\n      for (const { docIdx, tf } of postings) {\n        const docLen = this.docLens[docIdx];\n        // BM25 公式\n        const tfNorm = (tf * (BM25_K1 + 1)) / (tf + BM25_K1 * (1 - BM25_B + BM25_B * docLen / this.avgDocLen));\n        scores[docIdx] += idf * tfNorm;\n      }\n    }\n\n    // 排序取 Top-K\n    const results = [];\n    for (let i = 0; i < this.docCount; i++) {\n      if (scores[i] > 0) {\n        results.push({ id: this.docs[i].id, score: scores[i] });\n      }\n    }\n    results.sort((a, b) => b.score - a.score);\n    return results.slice(0, topK);\n  }\n}\n\n/**\n * 调用 Embedding API（异步，使用 fetch）\n * @param {string} text - 要嵌入的文本\n * @param {object} config - {provider, apiKey, baseUrl, model}\n * @returns {Promise<Float32Array|null>} 嵌入向量\n */\nasync function callEmbeddingAPI(text, config) {\n  const url = config.baseUrl || getEmbeddingUrl(config.provider);\n  if (!url || !config.apiKey) return null;\n\n  const model = config.model || EMBEDDING_MODELS[config.provider] || 'deepseek-embedding';\n\n  try {\n    const resp = await fetch(url, {\n      method: 'POST',\n      headers: {\n        'Content-Type': 'application/json',\n        'Authorization': `Bearer ${config.apiKey}`,\n      },\n      body: JSON.stringify({\n        model,\n        input: text.slice(0, 8000),\n      }),\n      signal: AbortSignal.timeout(30000),\n    });\n\n    if (!resp.ok) {\n      console.warn('Embedding API HTTP 错误:', resp.status);\n      return null;\n    }\n\n    const json = await resp.json();\n    if (json.data && json.data[0] && json.data[0].embedding) {\n      return new Float32Array(json.data[0].embedding);\n    }\n    if (json.error) {\n      console.warn('Embedding API 错误:', json.error.message || json.error);\n    }\n    return null;\n  } catch (e) {\n    console.warn('Embedding API 调用失败:', e.message);\n    return null;\n  }\n}\n\n/**\n * 批量调用 Embedding API（异步）\n * @param {string[]} texts - 文本数组\n * @param {object} config - API 配置\n * @returns {Promise<Float32Array[]>} 嵌入向量数组（失败的用 null 占位）\n */\nasync function callEmbeddingBatch(texts, config) {\n  const url = config.baseUrl || getEmbeddingUrl(config.provider);\n  if (!url || !config.apiKey) return texts.map(() => null);\n\n  const model = config.model || EMBEDDING_MODELS[config.provider] || 'deepseek-embedding';\n  const results = new Array(texts.length).fill(null);\n\n  // 分批处理\n  for (let i = 0; i < texts.length; i += EMBEDDING_BATCH_SIZE) {\n    const batch = texts.slice(i, i + EMBEDDING_BATCH_SIZE);\n\n    try {\n      const resp = await fetch(url, {\n        method: 'POST',\n        headers: {\n          'Content-Type': 'application/json',\n          'Authorization': `Bearer ${config.apiKey}`,\n        },\n        body: JSON.stringify({\n          model,\n          input: batch.map(t => t.slice(0, 8000)),\n        }),\n        signal: AbortSignal.timeout(60000),\n      });\n\n      if (!resp.ok) {\n        console.warn('Embedding 批量 API HTTP 错误:', resp.status);\n        continue;\n      }\n\n      const json = await resp.json();\n      if (json.data) {\n        for (const item of json.data) {\n          results[i + item.index] = new Float32Array(item.embedding);\n        }\n      }\n      if (json.error) {\n        console.warn('Embedding 批量 API 错误:', json.error.message || json.error);\n      }\n    } catch (e) {\n      console.warn('Embedding 批量 API 调用失败:', e.message);\n    }\n\n    // 批次间隔，避免限流\n    if (i + EMBEDDING_BATCH_SIZE < texts.length) {\n      await new Promise(r => setTimeout(r, 1000));\n    }\n  }\n\n  return results;\n}\n\n/**\n * 从 settings 读取 Embedding 配置\n * 支持独立的 embedding 配置，回退到通用 ai 配置\n */\nfunction getEmbeddingConfig() {\n  const db = getDb();\n  const rows = db.prepare(\n    \"SELECT key, value FROM settings WHERE key IN ('embedding_provider','embedding_api_key','embedding_base_url','embedding_model','ai_provider','ai_api_key','ai_base_url')\"\n  ).all();\n  const config = {};\n  for (const row of rows) config[row.key] = row.value;\n  return {\n    provider: config.embedding_provider || config.ai_provider || 'deepseek',\n    apiKey: config.embedding_api_key || config.ai_api_key || '',\n    baseUrl: config.embedding_base_url || '',\n    model: config.embedding_model || '',\n  };\n}\n\n// ===== 向量索引（使用 sqlite-vec）=====\n// 向量表用自增整数 ID，chunkMeta 映射 vectorId → {source, parentId, chunkIndex}\nlet vectorIdCounter = 0;\nlet chunkMeta = {}; // vectorId → {source, parentId, chunkIndex, chunkId}\nlet chunkIdToVectorId = {}; // chunkId → vectorId\n\nclass VectorIndex {\n  constructor() {\n    this.dimension = EMBEDDING_DIMENSION;\n  }\n\n  initTable(db) {\n    ensureVec(db);\n    db.exec(`\n      CREATE VIRTUAL TABLE IF NOT EXISTS knowledge_vec USING vec0(\n        id INTEGER PRIMARY KEY,\n        embedding float[${this.dimension}]\n      );\n    `);\n  }\n\n  rebuildTable(db) {\n    ensureVec(db);\n    try {\n      const info = db.prepare(\"SELECT sql FROM sqlite_master WHERE name = 'knowledge_vec'\").get();\n      if (info && !info.sql.includes(`float[${this.dimension}]`)) {\n        console.log(`向量表维度变化，重建表 (${this.dimension}维)`);\n        db.prepare('DROP TABLE IF EXISTS knowledge_vec').run();\n      }\n    } catch {}\n    this.initTable(db);\n  }\n\n  upsert(db, vectorId, embedding) {\n    if (!embedding) return false;\n    try {\n      db.prepare('DELETE FROM knowledge_vec WHERE id = ?').run(vectorId);\n      db.prepare('INSERT INTO knowledge_vec (id, embedding) VALUES (?, ?)').run(\n        vectorId,\n        Buffer.from(embedding.buffer)\n      );\n      return true;\n    } catch (e) {\n      return false;\n    }\n  }\n\n  search(db, queryVector, topK = 5) {\n    ensureVec(db);\n    if (!queryVector) return [];\n    try {\n      const results = db.prepare(`\n        SELECT id, distance FROM knowledge_vec WHERE embedding MATCH ? ORDER BY distance LIMIT ?\n      `).all(Buffer.from(queryVector.buffer), topK);\n      return results.map(r => ({ vectorId: r.id, score: 1 - r.distance }));\n    } catch (e) {\n      return [];\n    }\n  }\n}\n\n// 全局实例\nlet bm25Index = null;\nconst vectorIndex = new VectorIndex();\n\n// ===== 文档拆分 (Chunking) =====\nconst CHUNK_MAX_CHARS = 400;\nconst CHUNK_OVERLAP = 50;\n\n/**\n * 将长文本按段落拆分成小块\n */\nfunction chunkText(text, maxChars = CHUNK_MAX_CHARS) {\n  if (!text || text.length <= maxChars) return [text];\n\n  const chunks = [];\n  const paragraphs = text.split(/\\n\\s*\\n/).filter(p => p.trim());\n\n  let current = '';\n  for (const para of paragraphs) {\n    if (current.length + para.length + 1 <= maxChars) {\n      current = current ? current + '\\n\\n' + para : para;\n    } else {\n      if (current) chunks.push(current);\n      if (para.length > maxChars) {\n        const sentences = para.split(/(?<=[。！？.!?\\n])/).filter(s => s.trim());\n        let sentBuf = '';\n        for (const sent of sentences) {\n          if (sentBuf.length + sent.length <= maxChars) {\n            sentBuf += sent;\n          } else {\n            if (sentBuf) chunks.push(sentBuf);\n            sentBuf = sent;\n          }\n        }\n        if (sentBuf) current = sentBuf;\n        else current = '';\n      } else {\n        current = para;\n      }\n    }\n  }\n  if (current) chunks.push(current);\n\n  if (CHUNK_OVERLAP > 0 && chunks.length > 1) {\n    const overlapped = [chunks[0]];\n    for (let i = 1; i < chunks.length; i++) {\n      const prevTail = chunks[i - 1].slice(-CHUNK_OVERLAP);\n      overlapped.push(prevTail + chunks[i]);\n    }\n    return overlapped;\n  }\n\n  return chunks;\n}\n\n/**\n * 将知识条目拆分成可索引的块\n * 使用字符串复合 ID: \"{source}:{parentId}:{chunkIndex}\" 避免命名空间冲突\n */\nfunction chunkItem(item) {\n  const fullText = item.content || '';\n  const ns = NS[item.source] || 'k';\n\n  if (fullText.length <= CHUNK_MAX_CHARS) {\n    return [{\n      chunkId: `${ns}:${item.id}:0`,\n      source: item.source,\n      parentId: item.id,\n      chunkIndex: 0,\n      title: item.title,\n      content: item.title + ' ' + fullText,\n      category: item.category,\n    }];\n  }\n\n  const chunks = chunkText(fullText);\n  return chunks.map((chunk, i) => ({\n    chunkId: `${ns}:${item.id}:${i}`,\n    source: item.source,\n    parentId: item.id,\n    chunkIndex: i,\n    title: item.title + (chunks.length > 1 ? ` (第${i + 1}/${chunks.length}部分)` : ''),\n    content: item.title + ' ' + chunk,\n    category: item.category,\n  }));\n}\n\n// 块内容缓存（chunkId → {title, content, category, source, parentId}）\nlet chunkContentCache = {};\n\n/**\n * 重建索引（异步）\n * @param {boolean} forceVectors - 强制重建向量\n * @returns {Promise<number>} 索引条目数\n */\nasync function rebuildIndex(forceVectors = false) {\n  const db = getDb();\n\n  // 1. ai_knowledge 表\n  const knowledgeItems = db.prepare(\n    \"SELECT id, title, content, category FROM ai_knowledge WHERE status = 'active'\"\n  ).all().map(item => ({ ...item, source: 'knowledge' }));\n\n  // 2. tutorials 表（已发布教程）\n  const tutorialItems = db.prepare(\n    \"SELECT id, title, content, category FROM tutorials WHERE status = 'published'\"\n  ).all().map(item => ({ ...item, source: 'tutorial' }));\n\n  // 3. faqs 表（激活状态）\n  const faqItems = db.prepare(\n    \"SELECT id, question as title, answer as content, category FROM faqs WHERE status = 'active'\"\n  ).all().map(item => ({ ...item, source: 'faq' }));\n\n  // 合并所有数据源\n  const rawItems = [...knowledgeItems, ...tutorialItems, ...faqItems];\n\n  // 拆分长文档\n  const allItems = [];\n  let chunkCount = 0;\n  for (const item of rawItems) {\n    const chunks = chunkItem(item);\n    allItems.push(...chunks);\n    if (chunks.length > 1) chunkCount++;\n  }\n  if (chunkCount > 0) {\n    console.log(`文档拆分: ${chunkCount} 篇长文档被拆分成 ${allItems.length} 个块`);\n  }\n\n  // 更新块内容缓存\n  chunkContentCache = {};\n  for (const item of allItems) {\n    chunkContentCache[item.chunkId] = {\n      title: item.title,\n      content: item.content,\n      category: item.category,\n      source: item.source,\n      parentId: item.parentId,\n    };\n  }\n\n  // 重建 BM25\n  bm25Index = new BM25Index();\n  bm25Index.build(allItems.map(item => ({ ...item, id: item.chunkId })));\n\n  // 重建向量（使用真实 Embedding API）\n  try {\n    vectorIndex.rebuildTable(db);\n\n    const currentVecCount = db.prepare('SELECT COUNT(*) as cnt FROM knowledge_vec').get().cnt;\n    if (!forceVectors && currentVecCount >= allItems.length) {\n      console.log(`向量索引已存在 (${currentVecCount} 条)，跳过重建`);\n    } else {\n      try { db.prepare('DELETE FROM knowledge_vec').run(); } catch {}\n\n      // 重置映射\n      vectorIdCounter = 0;\n      chunkMeta = {};\n      chunkIdToVectorId = {};\n\n      const texts = allItems.map(item => item.title + ' ' + item.content);\n      const embedConfig = getEmbeddingConfig();\n\n      if (embedConfig.apiKey) {\n        console.log(`正在调用 ${embedConfig.provider} Embedding API，共 ${texts.length} 条...`);\n        const embeddings = await callEmbeddingBatch(texts, embedConfig);\n\n        let successCount = 0;\n        for (let i = 0; i < allItems.length; i++) {\n          const vecId = ++vectorIdCounter;\n          if (embeddings[i] && vectorIndex.upsert(db, vecId, embeddings[i])) {\n            chunkMeta[vecId] = {\n              source: allItems[i].source,\n              parentId: allItems[i].parentId,\n              chunkIndex: allItems[i].chunkIndex,\n              chunkId: allItems[i].chunkId,\n            };\n            chunkIdToVectorId[allItems[i].chunkId] = vecId;\n            successCount++;\n          }\n        }\n        console.log(`向量索引已建立: ${successCount}/${allItems.length} 条成功`);\n      } else {\n        console.warn('未配置 AI API Key，跳过向量索引构建（仅使用 BM25）');\n      }\n    }\n  } catch (e) {\n    console.warn('向量索引初始化失败（不影响 BM25 检索）:', e.message);\n  }\n\n  console.log(`RAG 索引已重建: ${allItems.length} 条 (知识库${knowledgeItems.length} + 教程${tutorialItems.length} + FAQ${faqItems.length})`);\n  return allItems.length;\n}\n\n/**\n * 混合检索：BM25 + 向量，加权合并\n * @param {string} query - 用户问题\n * @param {number} topK - 返回条数\n * @returns {Promise<Array>} [{chunkId, title, content, score, source, parentId}]\n */\nasync function retrieve(query, topK = 5) {\n  const db = getDb();\n\n  if (!bm25Index) await rebuildIndex();\n\n  // BM25 检索\n  const bm25Results = bm25Index.search(query, topK * 2);\n\n  // 向量检索\n  let vecResults = [];\n  try {\n    const embedConfig = getEmbeddingConfig();\n    if (embedConfig.apiKey) {\n      const queryVector = await callEmbeddingAPI(query, embedConfig);\n      if (queryVector) {\n        const rawVecResults = vectorIndex.search(db, queryVector, topK * 2);\n        // 将 vectorId 映射回 chunkId\n        for (const r of rawVecResults) {\n          const meta = chunkMeta[r.vectorId];\n          if (meta) {\n            vecResults.push({ id: meta.chunkId, score: r.score });\n          }\n        }\n      }\n    }\n  } catch {}\n\n  // 合并分数（加权：BM25 0.6 + 向量 0.4）\n  const scoreMap = {};\n  for (const r of bm25Results) {\n    scoreMap[r.id] = (scoreMap[r.id] || 0) + r.score * 0.6;\n  }\n  for (const r of vecResults) {\n    scoreMap[r.id] = (scoreMap[r.id] || 0) + r.score * 0.4;\n  }\n\n  const sorted = Object.entries(scoreMap)\n    .map(([id, score]) => ({ chunkId: id, score }))\n    .sort((a, b) => b.score - a.score)\n    .slice(0, topK);\n\n  if (sorted.length === 0) return [];\n\n  // 获取详情\n  const itemMap = {};\n  const uncachedItems = [];\n\n  for (const s of sorted) {\n    if (chunkContentCache[s.chunkId]) {\n      itemMap[s.chunkId] = chunkContentCache[s.chunkId];\n    } else {\n      uncachedItems.push(s);\n    }\n  }\n\n  // 缓存未命中的，按来源查数据库\n  const knowledgeParents = new Set();\n  const tutorialParents = new Set();\n  const faqParents = new Set();\n\n  for (const s of uncachedItems) {\n    const [source, parentId] = s.chunkId.split(':');\n    const pid = Number(parentId);\n    if (source === 'k') knowledgeParents.add(pid);\n    else if (source === 't') tutorialParents.add(pid);\n    else if (source === 'f') faqParents.add(pid);\n  }\n\n  if (knowledgeParents.size > 0) {\n    const ids = [...knowledgeParents];\n    const ph = ids.map(() => '?').join(',');\n    const rows = db.prepare(`SELECT id, title, content, category FROM ai_knowledge WHERE id IN (${ph}) AND status = 'active'`).all(...ids);\n    for (const r of rows) {\n      itemMap[`k:${r.id}:0`] = { ...r, source: 'knowledge', parentId: r.id };\n    }\n  }\n  if (tutorialParents.size > 0) {\n    const ids = [...tutorialParents];\n    const ph = ids.map(() => '?').join(',');\n    const rows = db.prepare(`SELECT id, title, content, category FROM tutorials WHERE id IN (${ph}) AND status = 'published'`).all(...ids);\n    for (const r of rows) {\n      itemMap[`t:${r.id}:0`] = { ...r, source: 'tutorial', parentId: r.id };\n    }\n  }\n  if (faqParents.size > 0) {\n    const ids = [...faqParents];\n    const ph = ids.map(() => '?').join(',');\n    const rows = db.prepare(`SELECT id, question as title, answer as content, category FROM faqs WHERE id IN (${ph}) AND status = 'active'`).all(...ids);\n    for (const r of rows) {\n      itemMap[`f:${r.id}:0`] = { ...r, source: 'faq', parentId: r.id };\n    }\n  }\n\n  return sorted\n    .filter(s => itemMap[s.chunkId])\n    .map(s => {\n      const item = itemMap[s.chunkId];\n      const images = [];\n      const content = item.content || '';\n      const mdImgs = content.match(/!\\[.*?\\]\\((https?:\\/\\/[^)]+)\\)/g) || [];\n      for (const m of mdImgs) {\n        const url = m.match(/\\((https?:\\/\\/[^)]+)\\)/);\n        if (url) images.push(url[1]);\n      }\n      const htmlImgs = content.match(/<img[^>]+src=[\"'](https?:\\/\\/[^\"']+)[\"'][^>]*>/gi) || [];\n      for (const m of htmlImgs) {\n        const url = m.match(/src=[\"'](https?:\\/\\/[^\"']+)[\"']/i);\n        if (url) images.push(url[1]);\n      }\n      return {\n        title: item.title,\n        content: item.content,\n        category: item.category,\n        source: item.source,\n        parentId: item.parentId,\n        score: s.score,\n        images: [...new Set(images)],\n      };\n    });\n}\n\n/**\n * 计算两个文本的相似度\n */\nfunction similarity(textA, textB) {\n  const tokensA = new Set(tokenize(textA));\n  const tokensB = new Set(tokenize(textB));\n  if (tokensA.size === 0 || tokensB.size === 0) return 0;\n  let overlap = 0;\n  for (const t of tokensA) {\n    if (tokensB.has(t)) overlap++;\n  }\n  const union = new Set([...tokensA, ...tokensB]).size;\n  return union > 0 ? overlap / union : 0;\n}\n\n/**\n * 查找与给定问题相似的已有自动学习条目\n */\nasync function findSimilarLearned(question, threshold = 2.0) {\n  const db = getDb();\n\n  if (!bm25Index) await rebuildIndex();\n\n  const bm25Results = bm25Index.search(question, 10);\n\n  for (const result of bm25Results) {\n    if (result.score < threshold) break;\n\n    // chunkId 格式: \"k:{parentId}:{chunkIndex}\"\n    const parts = result.id.split(':');\n    const parentId = Number(parts[1]);\n\n    const item = db.prepare(\n      \"SELECT id, title, content FROM ai_knowledge WHERE id = ? AND category = '自动学习' AND status = 'active'\"\n    ).get(parentId);\n\n    if (item) {\n      return { ...item, score: result.score };\n    }\n  }\n\n  return null;\n}\n\n/**\n * 清理自动学习条目\n */\nfunction cleanupLearned(maxCount = 200) {\n  const db = getDb();\n  const items = db.prepare(\n    \"SELECT id, title, content FROM ai_knowledge WHERE category = '自动学习' AND status = 'active' ORDER BY updated_at DESC\"\n  ).all();\n\n  if (items.length === 0) return { merged: 0, removed: 0 };\n\n  let merged = 0;\n  const toDelete = new Set();\n\n  for (let i = 0; i < items.length; i++) {\n    if (toDelete.has(items[i].id)) continue;\n    for (let j = i + 1; j < items.length; j++) {\n      if (toDelete.has(items[j].id)) continue;\n      const score = similarity(items[i].title, items[j].title);\n      if (score >= 0.4) {\n        const existingContent = items[i].content;\n        const newAnswer = items[j].content.replace(/^用户问题：.*?\\n\\n参考回答：/s, '');\n        if (newAnswer && !existingContent.includes(newAnswer.slice(0, 50))) {\n          db.prepare('UPDATE ai_knowledge SET content = content || ? WHERE id = ?')\n            .run('\\n\\n补充回答：' + newAnswer, items[i].id);\n        }\n        toDelete.add(items[j].id);\n        merged++;\n      }\n    }\n  }\n\n  if (toDelete.size > 0) {\n    const placeholders = Array.from(toDelete).map(() => '?').join(',');\n    db.prepare(`DELETE FROM ai_knowledge WHERE id IN (${placeholders})`).run(...toDelete);\n  }\n\n  let removed = 0;\n  const remaining = db.prepare(\n    \"SELECT COUNT(*) as cnt FROM ai_knowledge WHERE category = '自动学习' AND status = 'active'\"\n  ).get().cnt;\n\n  if (remaining > maxCount) {\n    const excess = remaining - maxCount;\n    const oldItems = db.prepare(\n      \"SELECT id FROM ai_knowledge WHERE category = '自动学习' AND status = 'active' ORDER BY updated_at ASC LIMIT ?\"\n    ).all(excess);\n    if (oldItems.length > 0) {\n      const placeholders = oldItems.map(() => '?').join(',');\n      db.prepare(`DELETE FROM ai_knowledge WHERE id IN (${placeholders})`).run(...oldItems.map(i => i.id));\n      removed = oldItems.length;\n    }\n  }\n\n  return { merged, removed };\n}\n\nmodule.exports = { retrieve, rebuildIndex, tokenize, similarity, findSimilarLearned, cleanupLearned };\n"}, "3434aeeecff5dee82ac201362a02222b99c273e2": {"t": 1785766730.3266647, "n": 21042, "h": "f42e1e0875ac35f9", "c": "/**\n * RAG 检索服务 — 混合检索（BM25 + 向量）+ Re-ranking\n * 向量检索使用真实 Embedding API（DeepSeek / OpenAI 兼容）\n */\nconst { getDb } = require('../database/schema');\nconst { execSync } = require('child_process');\n\nlet sqliteVecLoaded = false;\nfunction ensureVec(db) {\n  if (sqliteVecLoaded) return;\n  try {\n    const sqliteVec = require('sqlite-vec');\n    sqliteVec.load(db);\n    sqliteVecLoaded = true;\n  } catch (e) {\n    console.warn('sqlite-vec 加载失败:', e.message);\n  }\n}\n\n// Embedding 配置\nconst EMBEDDING_DIMENSION = 512; // bge-small-zh-v1.5 输出维度\nconst EMBEDDING_BATCH_SIZE = 10; // DashScope 限制每批最多 10 条\n\n// Embedding 模型映射\nconst EMBEDDING_MODELS = {\n  deepseek: 'deepseek-embedding',\n  openai: 'text-embedding-3-small',\n  qwen: 'text-embedding-v3',\n};\n\n// Provider 对应的 Embedding API URL\nfunction getEmbeddingUrl(provider) {\n  const urls = {\n    deepseek: 'https://api.deepseek.com/embeddings',\n    openai: 'https://api.openai.com/v1/embeddings',\n    qwen: 'https://dashscope.aliyuncs.com/compatible-mode/v1/embeddings',\n  };\n  return urls[provider] || '';\n}\n\n// BM25 参数\nconst BM25_K1 = 1.5;\nconst BM25_B = 0.75;\n\n// 常用同义词扩展\nconst SYNONYMS = {\n  '收费': ['价格', '多少钱', '费用', '售价', '报价'],\n  '价格': ['收费', '多少钱', '费用', '售价'],\n  '多少钱': ['价格', '收费', '费用'],\n  '购买': ['买', '下单', '订购'],\n  '登录': ['登陆', '登入', '进入后台'],\n  '系统': ['平台', '后台', '软件'],\n  '手机': ['设备', 'AI手机'],\n  '更新': ['升级', '版本更新'],\n  '问题': ['故障', 'bug', '报错', '出错'],\n};\n\n// 中文分词（bigram + 同义词扩展）\nfunction tokenize(text) {\n  if (!text) return [];\n  const cleaned = text.replace(/[^\\u4e00-\\u9fa5a-zA-Z0-9]/g, ' ').toLowerCase();\n  const tokens = [];\n  // 英文单词\n  const englishWords = cleaned.match(/[a-z0-9]+/g) || [];\n  tokens.push(...englishWords);\n  // 中文：bigram + trigram\n  const chinese = cleaned.replace(/[^\\u4e00-\\u9fa5]/g, '');\n  for (let i = 0; i < chinese.length - 1; i++) {\n    tokens.push(chinese.slice(i, i + 2));\n  }\n  for (let i = 0; i < chinese.length - 2; i++) {\n    tokens.push(chinese.slice(i, i + 3));\n  }\n  // 单字\n  for (const char of chinese) {\n    tokens.push(char);\n  }\n  // 同义词扩展\n  const expanded = [];\n  for (const token of tokens) {\n    expanded.push(token);\n    if (SYNONYMS[token]) {\n      expanded.push(...SYNONYMS[token]);\n    }\n  }\n  return expanded.filter(t => t.length > 0);\n}\n\n// BM25 索引\nclass BM25Index {\n  constructor() {\n    this.docs = []; // {id, tokens, length}\n    this.df = {};   // document frequency\n    this.totalDocs = 0;\n    this.avgDocLen = 0;\n  }\n\n  build(items) {\n    this.docs = [];\n    this.df = {};\n    this.totalDocs = items.length;\n    let totalLen = 0;\n\n    for (const item of items) {\n      const tokens = tokenize(item.title + ' ' + item.content);\n      this.docs.push({ id: item.id, tokens, length: tokens.length });\n      totalLen += tokens.length;\n\n      const uniqueTokens = new Set(tokens);\n      for (const t of uniqueTokens) {\n        this.df[t] = (this.df[t] || 0) + 1;\n      }\n    }\n    this.avgDocLen = this.docs.reduce((sum, d) => sum + d.length, 0) / this.totalDocs;\n  }\n\n  search(query, topK = 5) {\n    const queryTokens = tokenize(query);\n    if (queryTokens.length === 0) return [];\n\n    const scores = this.docs.map(doc => {\n      let score = 0;\n      const tf = {};\n      for (const t of doc.tokens) tf[t] = (tf[t] || 0) + 1;\n\n      for (const qt of queryTokens) {\n        if (!tf[qt]) continue;\n        const termFreq = tf[qt];\n        const docFreq = this.df[qt] || 0;\n        const idf = Math.log((this.totalDocs - docFreq + 0.5) / (docFreq + 0.5) + 1);\n        const tfNorm = (termFreq * (BM25_K1 + 1)) / (termFreq + BM25_K1 * (1 - BM25_B + BM25_B * doc.length / this.avgDocLen));\n        score += idf * tfNorm;\n      }\n      return { id: doc.id, score };\n    });\n\n    return scores.sort((a, b) => b.score - a.score).slice(0, topK);\n  }\n}\n\n/**\n * 调用 Embedding API（同步，通过 curl）\n * @param {string} text - 要嵌入的文本\n * @param {object} config - {provider, apiKey, baseUrl, model}\n * @returns {Float32Array|null} 嵌入向量\n */\nfunction callEmbeddingAPI(text, config) {\n  const url = config.baseUrl || getEmbeddingUrl(config.provider);\n  if (!url || !config.apiKey) return null;\n\n  const model = config.model || EMBEDDING_MODELS[config.provider] || 'deepseek-embedding';\n  const body = JSON.stringify({\n    model,\n    input: text.slice(0, 8000), // 截断超长文本\n  });\n\n  try {\n    const result = execSync(`curl -s -X POST \"${url}\" \\\n      -H \"Content-Type: application/json\" \\\n      -H \"Authorization: Bearer ${config.apiKey}\" \\\n      -d '${body.replace(/'/g, \"'\\''\")}'`, {\n      encoding: 'utf-8',\n      timeout: 30000,\n    });\n\n    const json = JSON.parse(result);\n    if (json.data && json.data[0] && json.data[0].embedding) {\n      return new Float32Array(json.data[0].embedding);\n    }\n    if (json.error) {\n      console.warn('Embedding API 错误:', json.error.message || json.error);\n    }\n    return null;\n  } catch (e) {\n    console.warn('Embedding API 调用失败:', e.message);\n    return null;\n  }\n}\n\n/**\n * 批量调用 Embedding API\n * @param {string[]} texts - 文本数组\n * @param {object} config - API 配置\n * @returns {Float32Array[]} 嵌入向量数组（失败的用 null 占位）\n */\nfunction callEmbeddingBatch(texts, config) {\n  const url = config.baseUrl || getEmbeddingUrl(config.provider);\n  if (!url || !config.apiKey) return texts.map(() => null);\n\n  const model = config.model || EMBEDDING_MODELS[config.provider] || 'deepseek-embedding';\n  const results = new Array(texts.length).fill(null);\n\n  // 分批处理\n  for (let i = 0; i < texts.length; i += EMBEDDING_BATCH_SIZE) {\n    const batch = texts.slice(i, i + EMBEDDING_BATCH_SIZE);\n    const body = JSON.stringify({\n      model,\n      input: batch.map(t => t.slice(0, 8000)),\n    });\n\n    try {\n      const result = execSync(`curl -s -X POST \"${url}\" \\\n        -H \"Content-Type: application/json\" \\\n        -H \"Authorization: Bearer ${config.apiKey}\" \\\n        -d '${body.replace(/'/g, \"'\\''\")}'`, {\n        encoding: 'utf-8',\n        timeout: 60000,\n      });\n\n      const json = JSON.parse(result);\n      if (json.data) {\n        // 按 index 排序并填充结果\n        for (const item of json.data) {\n          results[i + item.index] = new Float32Array(item.embedding);\n        }\n      }\n      if (json.error) {\n        console.warn('Embedding 批量 API 错误:', json.error.message || json.error);\n      }\n    } catch (e) {\n      console.warn('Embedding 批量 API 调用失败:', e.message);\n    }\n\n    // 批次间隔，避免限流\n    if (i + EMBEDDING_BATCH_SIZE < texts.length) {\n      try { execSync('sleep 1'); } catch {}\n    }\n  }\n\n  return results;\n}\n\n/**\n * 从 settings 读取 Embedding 配置\n * 支持独立的 embedding 配置，回退到通用 ai 配置\n */\nfunction getEmbeddingConfig() {\n  const db = getDb();\n  const rows = db.prepare(\n    \"SELECT key, value FROM settings WHERE key IN ('embedding_provider','embedding_api_key','embedding_base_url','embedding_model','ai_provider','ai_api_key','ai_base_url')\"\n  ).all();\n  const config = {};\n  for (const row of rows) config[row.key] = row.value;\n  return {\n    provider: config.embedding_provider || config.ai_provider || 'deepseek',\n    apiKey: config.embedding_api_key || config.ai_api_key || '',\n    baseUrl: config.embedding_base_url || '',\n    model: config.embedding_model || '',\n  };\n}\n\n// 向量索引（使用真实 Embedding API）\nclass VectorIndex {\n  constructor() {\n    this.dimension = EMBEDDING_DIMENSION;\n  }\n\n  initTable(db) {\n    ensureVec(db);\n    db.exec(`\n      CREATE VIRTUAL TABLE IF NOT EXISTS knowledge_vec USING vec0(\n        id INTEGER PRIMARY KEY,\n        embedding float[${this.dimension}]\n      );\n    `);\n  }\n\n  /**\n   * 重建向量表（维度变化时删除重建）\n   */\n  rebuildTable(db) {\n    ensureVec(db);\n    try {\n      // 检查现有表的维度\n      const info = db.prepare(\"SELECT sql FROM sqlite_master WHERE name = 'knowledge_vec'\").get();\n      if (info && !info.sql.includes(`float[${this.dimension}]`)) {\n        console.log(`向量表维度变化，重建表 (${this.dimension}维)`);\n        db.prepare('DROP TABLE IF EXISTS knowledge_vec').run();\n      }\n    } catch {}\n    this.initTable(db);\n  }\n\n  upsert(db, id, text, embedding) {\n    if (!embedding) return false;\n    const intId = Math.floor(Number(id));\n    if (isNaN(intId) || intId <= 0) return false;\n    try {\n      db.prepare('DELETE FROM knowledge_vec WHERE id = ' + intId).run();\n      db.prepare('INSERT INTO knowledge_vec (id, embedding) VALUES (' + intId + ', ?)').run(\n        Buffer.from(embedding.buffer)\n      );\n      return true;\n    } catch (e) {\n      return false;\n    }\n  }\n\n  search(db, queryVector, topK = 5) {\n    ensureVec(db);\n    if (!queryVector) return [];\n    try {\n      const results = db.prepare(`\n        SELECT id, distance FROM knowledge_vec WHERE embedding MATCH ? ORDER BY distance LIMIT ?\n      `).all(Buffer.from(queryVector.buffer), topK);\n      return results.map(r => ({ id: r.id, score: 1 - r.distance }));\n    } catch (e) {\n      return [];\n    }\n  }\n}\n\n// 全局实例\nlet bm25Index = null;\nconst vectorIndex = new VectorIndex();\n\n// ===== 文档拆分 (Chunking) =====\nconst CHUNK_MAX_CHARS = 400; // 每块最大字符数\nconst CHUNK_OVERLAP = 50;    // 块之间重叠字符数\n\n/**\n * 将长文本按段落拆分成小块\n * 优先按段落拆分，段落太长则按句子拆分\n * @param {string} text - 原始文本\n * @param {number} maxChars - 每块最大字符数\n * @returns {string[]} 拆分后的文本块\n */\nfunction chunkText(text, maxChars = CHUNK_MAX_CHARS) {\n  if (!text || text.length <= maxChars) return [text];\n\n  const chunks = [];\n  // 按段落拆分（双换行）\n  const paragraphs = text.split(/\\n\\s*\\n/).filter(p => p.trim());\n\n  let current = '';\n  for (const para of paragraphs) {\n    if (current.length + para.length + 1 <= maxChars) {\n      current = current ? current + '\\n\\n' + para : para;\n    } else {\n      if (current) chunks.push(current);\n      // 单个段落超长，按句子拆分\n      if (para.length > maxChars) {\n        const sentences = para.split(/(?<=[。！？.!?\\n])/).filter(s => s.trim());\n        let sentBuf = '';\n        for (const sent of sentences) {\n          if (sentBuf.length + sent.length <= maxChars) {\n            sentBuf += sent;\n          } else {\n            if (sentBuf) chunks.push(sentBuf);\n            sentBuf = sent;\n          }\n        }\n        if (sentBuf) current = sentBuf;\n        else current = '';\n      } else {\n        current = para;\n      }\n    }\n  }\n  if (current) chunks.push(current);\n\n  // 添加重叠：每块末尾保留 overlap 字符到下一块开头\n  if (CHUNK_OVERLAP > 0 && chunks.length > 1) {\n    const overlapped = [chunks[0]];\n    for (let i = 1; i < chunks.length; i++) {\n      const prevTail = chunks[i - 1].slice(-CHUNK_OVERLAP);\n      overlapped.push(prevTail + chunks[i]);\n    }\n    return overlapped;\n  }\n\n  return chunks;\n}\n\n/**\n * 将知识条目拆分成可索引的块\n * 短文档不拆分，长文档按段落拆分\n * @param {object} item - {id, title, content, category, source}\n * @returns {object[]} 拆分后的块数组 [{id, title, content, source}]\n */\nfunction chunkItem(item) {\n  const fullText = item.content || '';\n  if (fullText.length <= CHUNK_MAX_CHARS) {\n    // 短文档不拆分\n    return [{\n      id: item.id,\n      title: item.title,\n      content: item.title + ' ' + fullText,\n      source: item.source,\n      category: item.category,\n    }];\n  }\n\n  const chunks = chunkText(fullText);\n  return chunks.map((chunk, i) => ({\n    // 用 id * 1000 + chunkIndex 作为块 ID\n    id: item.id * 1000 + i,\n    title: item.title + (chunks.length > 1 ? ` (第${i + 1}/${chunks.length}部分)` : ''),\n    content: item.title + ' ' + chunk,\n    source: item.source,\n    category: item.category,\n    parentId: item.id,\n  }));\n}\n\n/**\n * 重建索引\n * 同时索引 ai_knowledge、tutorials（已发布）、faqs（激活状态）\n */\n// 块内容缓存（chunkId -> {title, content, category}）\nlet chunkContentCache = {};\n\n/**\n * 重建索引\n * @param {boolean} forceVectors - 强制重建向量（CRUD 操作后需要）\n */\nfunction rebuildIndex(forceVectors = false) {\n  const db = getDb();\n\n  // 1. ai_knowledge 表\n  const knowledgeItems = db.prepare(\n    \"SELECT id, title, content, category FROM ai_knowledge WHERE status = 'active'\"\n  ).all().map(item => ({\n    ...item,\n    source: 'knowledge',\n  }));\n\n  // 2. tutorials 表（已发布教程）\n  const tutorialItems = db.prepare(\n    \"SELECT id, title, content, category FROM tutorials WHERE status = 'published'\"\n  ).all().map(item => ({\n    ...item,\n    // 教程 id 加偏移避免与 ai_knowledge 冲突\n    id: 100000 + item.id,\n    source: 'tutorial',\n  }));\n\n  // 3. faqs 表（激活状态）\n  const faqItems = db.prepare(\n    \"SELECT id, question as title, answer as content, category FROM faqs WHERE status = 'active'\"\n  ).all().map(item => ({\n    ...item,\n    id: 200000 + item.id,\n    source: 'faq',\n  }));\n\n  // 合并所有数据源\n  const rawItems = [...knowledgeItems, ...tutorialItems, ...faqItems];\n\n  // 拆分长文档\n  const allItems = [];\n  let chunkCount = 0;\n  for (const item of rawItems) {\n    const chunks = chunkItem(item);\n    allItems.push(...chunks);\n    if (chunks.length > 1) chunkCount++;\n  }\n  if (chunkCount > 0) {\n    console.log(`文档拆分: ${chunkCount} 篇长文档被拆分成 ${allItems.length} 个块`);\n  }\n\n  // 更新块内容缓存\n  chunkContentCache = {};\n  for (const item of allItems) {\n    chunkContentCache[item.id] = {\n      title: item.title,\n      content: item.content,\n      category: item.category,\n    };\n  }\n\n  // 重建 BM25\n  bm25Index = new BM25Index();\n  bm25Index.build(allItems);\n\n  // 重建向量（使用真实 Embedding API）\n  try {\n    vectorIndex.rebuildTable(db);\n\n    // 检查是否需要重建向量（避免每次启动都调 API）\n    const currentVecCount = db.prepare('SELECT COUNT(*) as cnt FROM knowledge_vec').get().cnt;\n    if (!forceVectors && currentVecCount >= allItems.length) {\n      console.log(`向量索引已存在 (${currentVecCount} 条)，跳过重建`);\n    } else {\n      // 清空旧数据并重建\n      try { db.prepare('DELETE FROM knowledge_vec').run(); } catch {}\n\n      // 准备批量文本\n      const texts = allItems.map(item => item.title + ' ' + item.content);\n      const embedConfig = getEmbeddingConfig();\n\n      if (embedConfig.apiKey) {\n        console.log(`正在调用 ${embedConfig.provider} Embedding API，共 ${texts.length} 条...`);\n        const embeddings = callEmbeddingBatch(texts, embedConfig);\n\n        let successCount = 0;\n        for (let i = 0; i < allItems.length; i++) {\n          if (embeddings[i] && vectorIndex.upsert(db, allItems[i].id, allItems[i].title, embeddings[i])) {\n            successCount++;\n          }\n        }\n        console.log(`向量索引已建立: ${successCount}/${allItems.length} 条成功`);\n      } else {\n        console.warn('未配置 AI API Key，跳过向量索引构建（仅使用 BM25）');\n      }\n    }\n  } catch (e) {\n    console.warn('向量索引初始化失败（不影响 BM25 检索）:', e.message);\n  }\n\n  console.log(`RAG 索引已重建: ${allItems.length} 条 (知识库${knowledgeItems.length} + 教程${tutorialItems.length} + FAQ${faqItems.length})`);\n  return allItems.length;\n}\n\n/**\n * 混合检索：BM25 + 向量，加权合并\n * @param {string} query - 用户问题\n * @param {number} topK - 返回条数\n * @returns {Array} [{id, title, content, score}]\n */\nfunction retrieve(query, topK = 5) {\n  const db = getDb();\n\n  // 确保索引已建\n  if (!bm25Index) rebuildIndex();\n\n  // BM25 检索\n  const bm25Results = bm25Index.search(query, topK * 2);\n\n  // 向量检索（使用真实 Embedding API）\n  let vecResults = [];\n  try {\n    const embedConfig = getEmbeddingConfig();\n    if (embedConfig.apiKey) {\n      const queryVector = callEmbeddingAPI(query, embedConfig);\n      if (queryVector) {\n        vecResults = vectorIndex.search(db, queryVector, topK * 2);\n      }\n    }\n  } catch {}\n\n  // 合并分数（加权：BM25 0.6 + 向量 0.4）\n  const scoreMap = {};\n  for (const r of bm25Results) {\n    scoreMap[r.id] = (scoreMap[r.id] || 0) + r.score * 0.6;\n  }\n  for (const r of vecResults) {\n    scoreMap[r.id] = (scoreMap[r.id] || 0) + r.score * 0.4;\n  }\n\n  // 排序取 Top-K\n  const sorted = Object.entries(scoreMap)\n    .map(([id, score]) => ({ id: Number(id), score }))\n    .sort((a, b) => b.score - a.score)\n    .slice(0, topK);\n\n  // 获取知识条目详情\n  if (sorted.length === 0) return [];\n\n  const itemMap = {};\n\n  // 先从块缓存中查找\n  const uncachedIds = [];\n  for (const s of sorted) {\n    if (chunkContentCache[s.id]) {\n      itemMap[s.id] = chunkContentCache[s.id];\n    } else {\n      uncachedIds.push(s);\n    }\n  }\n\n  // 缓存未命中的，按 ID 范围查数据库\n  const knowledgeIds = uncachedIds.filter(s => s.id < 100000).map(s => s.id);\n  const tutorialIds = uncachedIds.filter(s => s.id >= 100000 && s.id < 200000).map(s => s.id - 100000);\n  const faqIds = uncachedIds.filter(s => s.id >= 200000).map(s => s.id - 200000);\n\n  if (knowledgeIds.length > 0) {\n    const ph = knowledgeIds.map(() => '?').join(',');\n    const rows = db.prepare(`SELECT id, title, content, category FROM ai_knowledge WHERE id IN (${ph}) AND status = 'active'`).all(...knowledgeIds);\n    for (const r of rows) itemMap[r.id] = r;\n  }\n  if (tutorialIds.length > 0) {\n    const ph = tutorialIds.map(() => '?').join(',');\n    const rows = db.prepare(`SELECT id, title, content, category FROM tutorials WHERE id IN (${ph}) AND status = 'published'`).all(...tutorialIds);\n    for (const r of rows) itemMap[100000 + r.id] = r;\n  }\n  if (faqIds.length > 0) {\n    const ph = faqIds.map(() => '?').join(',');\n    const rows = db.prepare(`SELECT id, question as title, answer as content, category FROM faqs WHERE id IN (${ph}) AND status = 'active'`).all(...faqIds);\n    for (const r of rows) itemMap[200000 + r.id] = r;\n  }\n\n  return sorted\n    .filter(s => itemMap[s.id])\n    .map(s => {\n      const item = itemMap[s.id];\n      // 从内容中提取图片 URL\n      const images = [];\n      const content = item.content || '';\n      // Markdown 图片: ![alt](url)\n      const mdImgs = content.match(/!\\[.*?\\]\\((https?:\\/\\/[^)]+)\\)/g) || [];\n      for (const m of mdImgs) {\n        const url = m.match(/\\((https?:\\/\\/[^)]+)\\)/);\n        if (url) images.push(url[1]);\n      }\n      // HTML img: <img src=\"url\">\n      const htmlImgs = content.match(/<img[^>]+src=[\"'](https?:\\/\\/[^\"']+)[\"'][^>]*>/gi) || [];\n      for (const m of htmlImgs) {\n        const url = m.match(/src=[\"'](https?:\\/\\/[^\"']+)[\"']/i);\n        if (url) images.push(url[1]);\n      }\n      return { ...item, score: s.score, images: [...new Set(images)] };\n    });\n}\n\n/**\n * 计算两个文本的相似度（基于 token 重叠率）\n * @returns {number} 0-1 之间的相似度\n */\nfunction similarity(textA, textB) {\n  const tokensA = new Set(tokenize(textA));\n  const tokensB = new Set(tokenize(textB));\n  if (tokensA.size === 0 || tokensB.size === 0) return 0;\n  let overlap = 0;\n  for (const t of tokensA) {\n    if (tokensB.has(t)) overlap++;\n  }\n  // Jaccard 相似度\n  const union = new Set([...tokensA, ...tokensB]).size;\n  return union > 0 ? overlap / union : 0;\n}\n\n/**\n * 查找与给定问题相似的已有自动学习条目（使用语义匹配，非 token 匹配）\n * @param {string} question - 用户问题\n * @param {number} threshold - 相似度阈值（默认 2.0，BM25 分数）\n * @returns {object|null} 相似条目或 null\n */\nfunction findSimilarLearned(question, threshold = 2.0) {\n  const db = getDb();\n\n  // 确保索引已建\n  if (!bm25Index) rebuildIndex();\n\n  // 用 BM25 语义检索（已包含同义词扩展）\n  const bm25Results = bm25Index.search(question, 10);\n\n  // 过滤出自动学习的条目\n  for (const result of bm25Results) {\n    if (result.score < threshold) break;\n\n    // 块 ID -> 原始 ID\n    const parentId = result.id >= 100000000 ? Math.floor(result.id / 1000) : result.id;\n\n    // 查数据库确认是自动学习条目\n    const item = db.prepare(\n      \"SELECT id, title, content FROM ai_knowledge WHERE id = ? AND category = '自动学习' AND status = 'active'\"\n    ).get(parentId);\n\n    if (item) {\n      return { ...item, score: result.score };\n    }\n  }\n\n  return null;\n}\n\n/**\n * 清理自动学习条目：合并相似内容，淘汰超限条目\n * @param {number} maxCount - 最大保留条数（默认 200）\n */\nfunction cleanupLearned(maxCount = 200) {\n  const db = getDb();\n  const items = db.prepare(\n    \"SELECT id, title, content FROM ai_knowledge WHERE category = '自动学习' AND status = 'active' ORDER BY updated_at DESC\"\n  ).all();\n\n  if (items.length === 0) return { merged: 0, removed: 0 };\n\n  let merged = 0;\n  const toDelete = new Set();\n\n  // 第一步：合并相似条目\n  for (let i = 0; i < items.length; i++) {\n    if (toDelete.has(items[i].id)) continue;\n    for (let j = i + 1; j < items.length; j++) {\n      if (toDelete.has(items[j].id)) continue;\n      const score = similarity(items[i].title, items[j].title);\n      if (score >= 0.4) {\n        // 把 j 的回答追加到 i，删除 j\n        const existingContent = items[i].content;\n        const newAnswer = items[j].content.replace(/^用户问题：.*?\\n\\n参考回答：/s, '');\n        if (newAnswer && !existingContent.includes(newAnswer.slice(0, 50))) {\n          db.prepare('UPDATE ai_knowledge SET content = content || ? WHERE id = ?')\n            .run('\\n\\n补充回答：' + newAnswer, items[i].id);\n        }\n        toDelete.add(items[j].id);\n        merged++;\n      }\n    }\n  }\n\n  // 批量删除被合并的条目\n  if (toDelete.size > 0) {\n    const placeholders = Array.from(toDelete).map(() => '?').join(',');\n    db.prepare(`DELETE FROM ai_knowledge WHERE id IN (${placeholders})`).run(...toDelete);\n  }\n\n  // 第二步：超限淘汰（删除最旧的）\n  let removed = 0;\n  const remaining = db.prepare(\n    \"SELECT COUNT(*) as cnt FROM ai_knowledge WHERE category = '自动学习' AND status = 'active'\"\n  ).get().cnt;\n\n  if (remaining > maxCount) {\n    const excess = remaining - maxCount;\n    const oldItems = db.prepare(\n      \"SELECT id FROM ai_knowledge WHERE category = '自动学习' AND status = 'active' ORDER BY updated_at ASC LIMIT ?\"\n    ).all(excess);\n    if (oldItems.length > 0) {\n      const placeholders = oldItems.map(() => '?').join(',');\n      db.prepare(`DELETE FROM ai_knowledge WHERE id IN (${placeholders})`).run(...oldItems.map(i => i.id));\n      removed = oldItems.length;\n    }\n  }\n\n  return { merged, removed };\n}\n\nmodule.exports = { retrieve, rebuildIndex, tokenize, similarity, findSimilarLearned, cleanupLearned };\n"}, "5d69dda26605c9190e699ef758b2b9224a7b823b": {"t": 1785766731.6512063, "n": 12312, "h": "d1f345c3503df612", "c": "/**\n * RAG 检索服务 — 混合检索（BM25 + 向量）+ Re-ranking\n */\nconst { getDb } = require('../database/schema');\nlet sqliteVecLoaded = false;\nfunction ensureVec(db) {\n  if (sqliteVecLoaded) return;\n  try {\n    const sqliteVec = require('sqlite-vec');\n    sqliteVec.load(db);\n    sqliteVecLoaded = true;\n  } catch (e) {\n    console.warn('sqlite-vec 加载失败:', e.message);\n  }\n}\n\n// BM25 参数\nconst BM25_K1 = 1.5;\nconst BM25_B = 0.75;\n\n// 常用同义词扩展\nconst SYNONYMS = {\n  '收费': ['价格', '多少钱', '费用', '售价', '报价'],\n  '价格': ['收费', '多少钱', '费用', '售价'],\n  '多少钱': ['价格', '收费', '费用'],\n  '购买': ['买', '下单', '订购'],\n  '登录': ['登陆', '登入', '进入后台'],\n  '系统': ['平台', '后台', '软件'],\n  '手机': ['设备', 'AI手机'],\n  '更新': ['升级', '版本更新'],\n  '问题': ['故障', 'bug', '报错', '出错'],\n};\n\n// 中文分词（bigram + 同义词扩展）\nfunction tokenize(text) {\n  if (!text) return [];\n  const cleaned = text.replace(/[^\\u4e00-\\u9fa5a-zA-Z0-9]/g, ' ').toLowerCase();\n  const tokens = [];\n  // 英文单词\n  const englishWords = cleaned.match(/[a-z0-9]+/g) || [];\n  tokens.push(...englishWords);\n  // 中文：bigram + trigram\n  const chinese = cleaned.replace(/[^\\u4e00-\\u9fa5]/g, '');\n  for (let i = 0; i < chinese.length - 1; i++) {\n    tokens.push(chinese.slice(i, i + 2));\n  }\n  for (let i = 0; i < chinese.length - 2; i++) {\n    tokens.push(chinese.slice(i, i + 3));\n  }\n  // 单字\n  for (const char of chinese) {\n    tokens.push(char);\n  }\n  // 同义词扩展\n  const expanded = [];\n  for (const token of tokens) {\n    expanded.push(token);\n    if (SYNONYMS[token]) {\n      expanded.push(...SYNONYMS[token]);\n    }\n  }\n  return expanded.filter(t => t.length > 0);\n}\n\n// BM25 索引\nclass BM25Index {\n  constructor() {\n    this.docs = []; // {id, tokens, length}\n    this.df = {};   // document frequency\n    this.totalDocs = 0;\n    this.avgDocLen = 0;\n  }\n\n  build(items) {\n    this.docs = [];\n    this.df = {};\n    this.totalDocs = items.length;\n    let totalLen = 0;\n\n    for (const item of items) {\n      const tokens = tokenize(item.title + ' ' + item.content);\n      this.docs.push({ id: item.id, tokens, length: tokens.length });\n      totalLen += tokens.length;\n\n      const uniqueTokens = new Set(tokens);\n      for (const t of uniqueTokens) {\n        this.df[t] = (this.df[t] || 0) + 1;\n      }\n    }\n    this.avgDocLen = this.docs.reduce((sum, d) => sum + d.length, 0) / this.totalDocs;\n  }\n\n  search(query, topK = 5) {\n    const queryTokens = tokenize(query);\n    if (queryTokens.length === 0) return [];\n\n    const scores = this.docs.map(doc => {\n      let score = 0;\n      const tf = {};\n      for (const t of doc.tokens) tf[t] = (tf[t] || 0) + 1;\n\n      for (const qt of queryTokens) {\n        if (!tf[qt]) continue;\n        const termFreq = tf[qt];\n        const docFreq = this.df[qt] || 0;\n        const idf = Math.log((this.totalDocs - docFreq + 0.5) / (docFreq + 0.5) + 1);\n        const tfNorm = (termFreq * (BM25_K1 + 1)) / (termFreq + BM25_K1 * (1 - BM25_B + BM25_B * doc.length / this.avgDocLen));\n        score += idf * tfNorm;\n      }\n      return { id: doc.id, score };\n    });\n\n    return scores.sort((a, b) => b.score - a.score).slice(0, topK);\n  }\n}\n\n// 向量索引（使用 sqlite-vec）\nclass VectorIndex {\n  constructor() {\n    this.dimension = 128; // 特征哈希维度\n  }\n\n  /**\n   * 简单特征哈希：将文本转为固定维度向量\n   * 用 SimHash 思想，对中文和英文都有效\n   */\n  textToVector(text) {\n    const tokens = tokenize(text);\n    const vector = new Float32Array(this.dimension);\n    for (let i = 0; i < tokens.length; i++) {\n      const token = tokens[i];\n      // 简单哈希\n      let hash = 0;\n      for (let j = 0; j < token.length; j++) {\n        hash = ((hash << 5) - hash + token.charCodeAt(j)) | 0;\n      }\n      const idx = Math.abs(hash) % this.dimension;\n      // 使用 +/-1 投影（特征哈希）\n      vector[idx] += (hash > 0 ? 1 : -1);\n    }\n    // L2 归一化\n    let norm = 0;\n    for (let i = 0; i < this.dimension; i++) norm += vector[i] * vector[i];\n    norm = Math.sqrt(norm) || 1;\n    for (let i = 0; i < this.dimension; i++) vector[i] /= norm;\n    return vector;\n  }\n\n  initTable(db) {\n    ensureVec(db);\n    db.exec(`\n      CREATE VIRTUAL TABLE IF NOT EXISTS knowledge_vec USING vec0(\n        id INTEGER PRIMARY KEY,\n        embedding float[${this.dimension}]\n      );\n    `);\n  }\n\n  upsert(db, id, text) {\n    const vector = this.textToVector(text);\n    const intId = Math.floor(Number(id));\n    if (isNaN(intId) || intId <= 0) return;\n    // 先删除旧的\n    try { db.prepare('DELETE FROM knowledge_vec WHERE id = ' + intId).run(); } catch {}\n    // 插入新的（sqlite-vec 要求主键必须是字面量整数，不能用参数绑定）\n    db.prepare('INSERT INTO knowledge_vec (id, embedding) VALUES (' + intId + ', ?)').run(Buffer.from(vector.buffer));\n  }\n\n  search(db, queryText, topK = 5) {\n    ensureVec(db);\n    const queryVector = this.textToVector(queryText);\n    try {\n      const results = db.prepare(`\n        SELECT id, distance FROM knowledge_vec WHERE embedding MATCH ? ORDER BY distance LIMIT ?\n      `).all(Buffer.from(queryVector.buffer), topK);\n      return results.map(r => ({ id: r.id, score: 1 - r.distance })); // distance -> similarity\n    } catch (e) {\n      // 如果向量表还没建好，返回空\n      return [];\n    }\n  }\n}\n\n// 全局实例\nlet bm25Index = null;\nconst vectorIndex = new VectorIndex();\n\n/**\n * 重建索引\n * 同时索引 ai_knowledge、tutorials（已发布）、faqs（激活状态）\n */\nfunction rebuildIndex() {\n  const db = getDb();\n\n  // 1. ai_knowledge 表\n  const knowledgeItems = db.prepare(\n    \"SELECT id, title, content, category FROM ai_knowledge WHERE status = 'active'\"\n  ).all().map(item => ({\n    ...item,\n    source: 'knowledge',\n  }));\n\n  // 2. tutorials 表（已发布教程）\n  const tutorialItems = db.prepare(\n    \"SELECT id, title, content, category FROM tutorials WHERE status = 'published'\"\n  ).all().map(item => ({\n    ...item,\n    // 教程 id 加偏移避免与 ai_knowledge 冲突\n    id: 100000 + item.id,\n    source: 'tutorial',\n  }));\n\n  // 3. faqs 表（激活状态）\n  const faqItems = db.prepare(\n    \"SELECT id, question as title, answer as content, category FROM faqs WHERE status = 'active'\"\n  ).all().map(item => ({\n    ...item,\n    id: 200000 + item.id,\n    source: 'faq',\n  }));\n\n  // 合并所有数据源\n  const allItems = [...knowledgeItems, ...tutorialItems, ...faqItems];\n\n  // 重建 BM25\n  bm25Index = new BM25Index();\n  bm25Index.build(allItems);\n\n  // 重建向量\n  try {\n    vectorIndex.initTable(db);\n    // 清空旧数据\n    try { db.prepare('DELETE FROM knowledge_vec').run(); } catch {}\n    let vecCount = 0;\n    for (const item of allItems) {\n      try {\n        vectorIndex.upsert(db, item.id, item.title + ' ' + item.content);\n        vecCount++;\n      } catch (e) {\n        // 单条失败不影响其他条目\n      }\n    }\n    if (vecCount < allItems.length) {\n      console.warn('向量索引: ' + (allItems.length - vecCount) + ' 条插入失败，已跳过');\n    }\n  } catch (e) {\n    console.warn('向量索引初始化失败（不影响 BM25 检索）:', e.message);\n  }\n\n  console.log(`RAG 索引已重建: ${allItems.length} 条 (知识库${knowledgeItems.length} + 教程${tutorialItems.length} + FAQ${faqItems.length})`);\n  return allItems.length;\n}\n\n/**\n * 混合检索：BM25 + 向量，加权合并\n * @param {string} query - 用户问题\n * @param {number} topK - 返回条数\n * @returns {Array} [{id, title, content, score}]\n */\nfunction retrieve(query, topK = 5) {\n  const db = getDb();\n\n  // 确保索引已建\n  if (!bm25Index) rebuildIndex();\n\n  // BM25 检索\n  const bm25Results = bm25Index.search(query, topK * 2);\n\n  // 向量检索\n  let vecResults = [];\n  try {\n    vecResults = vectorIndex.search(db, query, topK * 2);\n  } catch {}\n\n  // 合并分数（加权：BM25 0.6 + 向量 0.4）\n  const scoreMap = {};\n  for (const r of bm25Results) {\n    scoreMap[r.id] = (scoreMap[r.id] || 0) + r.score * 0.6;\n  }\n  for (const r of vecResults) {\n    scoreMap[r.id] = (scoreMap[r.id] || 0) + r.score * 0.4;\n  }\n\n  // 排序取 Top-K\n  const sorted = Object.entries(scoreMap)\n    .map(([id, score]) => ({ id: Number(id), score }))\n    .sort((a, b) => b.score - a.score)\n    .slice(0, topK);\n\n  // 获取知识条目详情（根据 ID 范围区分数据源）\n  if (sorted.length === 0) return [];\n\n  const itemMap = {};\n\n  // 按 ID 范围分组查询\n  const knowledgeIds = sorted.filter(s => s.id < 100000).map(s => s.id);\n  const tutorialIds = sorted.filter(s => s.id >= 100000 && s.id < 200000).map(s => s.id - 100000);\n  const faqIds = sorted.filter(s => s.id >= 200000).map(s => s.id - 200000);\n\n  if (knowledgeIds.length > 0) {\n    const ph = knowledgeIds.map(() => '?').join(',');\n    const rows = db.prepare(`SELECT id, title, content, category FROM ai_knowledge WHERE id IN (${ph}) AND status = 'active'`).all(...knowledgeIds);\n    for (const r of rows) itemMap[r.id] = r;\n  }\n  if (tutorialIds.length > 0) {\n    const ph = tutorialIds.map(() => '?').join(',');\n    const rows = db.prepare(`SELECT id, title, content, category FROM tutorials WHERE id IN (${ph}) AND status = 'published'`).all(...tutorialIds);\n    for (const r of rows) itemMap[100000 + r.id] = r;\n  }\n  if (faqIds.length > 0) {\n    const ph = faqIds.map(() => '?').join(',');\n    const rows = db.prepare(`SELECT id, question as title, answer as content, category FROM faqs WHERE id IN (${ph}) AND status = 'active'`).all(...faqIds);\n    for (const r of rows) itemMap[200000 + r.id] = r;\n  }\n\n  return sorted\n    .filter(s => itemMap[s.id])\n    .map(s => ({ ...itemMap[s.id], score: s.score }));\n}\n\n/**\n * 计算两个文本的相似度（基于 token 重叠率）\n * @returns {number} 0-1 之间的相似度\n */\nfunction similarity(textA, textB) {\n  const tokensA = new Set(tokenize(textA));\n  const tokensB = new Set(tokenize(textB));\n  if (tokensA.size === 0 || tokensB.size === 0) return 0;\n  let overlap = 0;\n  for (const t of tokensA) {\n    if (tokensB.has(t)) overlap++;\n  }\n  // Jaccard 相似度\n  const union = new Set([...tokensA, ...tokensB]).size;\n  return union > 0 ? overlap / union : 0;\n}\n\n/**\n * 查找与给定问题相似的已有自动学习条目\n * @param {string} question - 用户问题\n * @param {number} threshold - 相似度阈值（默认 0.35）\n * @returns {object|null} 相似条目或 null\n */\nfunction findSimilarLearned(question, threshold = 0.35) {\n  const db = getDb();\n  const existing = db.prepare(\n    \"SELECT id, title, content FROM ai_knowledge WHERE category = '自动学习' AND status = 'active'\"\n  ).all();\n\n  let bestMatch = null;\n  let bestScore = 0;\n\n  for (const item of existing) {\n    const score = similarity(question, item.title);\n    if (score > bestScore && score >= threshold) {\n      bestScore = score;\n      bestMatch = item;\n    }\n  }\n\n  return bestMatch ? { ...bestMatch, score: bestScore } : null;\n}\n\n/**\n * 清理自动学习条目：合并相似内容，淘汰超限条目\n * @param {number} maxCount - 最大保留条数（默认 200）\n */\nfunction cleanupLearned(maxCount = 200) {\n  const db = getDb();\n  const items = db.prepare(\n    \"SELECT id, title, content FROM ai_knowledge WHERE category = '自动学习' AND status = 'active' ORDER BY updated_at DESC\"\n  ).all();\n\n  if (items.length === 0) return { merged: 0, removed: 0 };\n\n  let merged = 0;\n  const toDelete = new Set();\n\n  // 第一步：合并相似条目\n  for (let i = 0; i < items.length; i++) {\n    if (toDelete.has(items[i].id)) continue;\n    for (let j = i + 1; j < items.length; j++) {\n      if (toDelete.has(items[j].id)) continue;\n      const score = similarity(items[i].title, items[j].title);\n      if (score >= 0.4) {\n        // 把 j 的回答追加到 i，删除 j\n        const existingContent = items[i].content;\n        const newAnswer = items[j].content.replace(/^用户问题：.*?\\n\\n参考回答：/s, '');\n        if (newAnswer && !existingContent.includes(newAnswer.slice(0, 50))) {\n          db.prepare('UPDATE ai_knowledge SET content = content || ? WHERE id = ?')\n            .run('\\n\\n补充回答：' + newAnswer, items[i].id);\n        }\n        toDelete.add(items[j].id);\n        merged++;\n      }\n    }\n  }\n\n  // 批量删除被合并的条目\n  if (toDelete.size > 0) {\n    const placeholders = Array.from(toDelete).map(() => '?').join(',');\n    db.prepare(`DELETE FROM ai_knowledge WHERE id IN (${placeholders})`).run(...toDelete);\n  }\n\n  // 第二步：超限淘汰（删除最旧的）\n  let removed = 0;\n  const remaining = db.prepare(\n    \"SELECT COUNT(*) as cnt FROM ai_knowledge WHERE category = '自动学习' AND status = 'active'\"\n  ).get().cnt;\n\n  if (remaining > maxCount) {\n    const excess = remaining - maxCount;\n    const oldItems = db.prepare(\n      \"SELECT id FROM ai_knowledge WHERE category = '自动学习' AND status = 'active' ORDER BY updated_at ASC LIMIT ?\"\n    ).all(excess);\n    if (oldItems.length > 0) {\n      const placeholders = oldItems.map(() => '?').join(',');\n      db.prepare(`DELETE FROM ai_knowledge WHERE id IN (${placeholders})`).run(...oldItems.map(i => i.id));\n      removed = oldItems.length;\n    }\n  }\n\n  return { merged, removed };\n}\n\nmodule.exports = { retrieve, rebuildIndex, tokenize, similarity, findSimilarLearned, cleanupLearned };\n"}, "f882ddb355bf9ea7f8eff6c1d5cd683e145d0aec": {"t": 1785766732.961888, "n": 7059, "h": "e27f50ac2c752407", "c": "/**\n * RAG 检索服务 — 混合检索（BM25 + 向量）+ Re-ranking\n */\nconst { getDb } = require('../database/schema');\nlet sqliteVecLoaded = false;\nfunction ensureVec(db) {\n  if (sqliteVecLoaded) return;\n  try {\n    const sqliteVec = require('sqlite-vec');\n    sqliteVec.load(db);\n    sqliteVecLoaded = true;\n  } catch (e) {\n    console.warn('sqlite-vec 加载失败:', e.message);\n  }\n}\n\n// BM25 参数\nconst BM25_K1 = 1.5;\nconst BM25_B = 0.75;\n\n// 常用同义词扩展\nconst SYNONYMS = {\n  '收费': ['价格', '多少钱', '费用', '售价', '报价'],\n  '价格': ['收费', '多少钱', '费用', '售价'],\n  '多少钱': ['价格', '收费', '费用'],\n  '购买': ['买', '下单', '订购'],\n  '登录': ['登陆', '登入', '进入后台'],\n  '系统': ['平台', '后台', '软件'],\n  '手机': ['设备', 'AI手机'],\n  '更新': ['升级', '版本更新'],\n  '问题': ['故障', 'bug', '报错', '出错'],\n};\n\n// 中文分词（bigram + 同义词扩展）\nfunction tokenize(text) {\n  if (!text) return [];\n  const cleaned = text.replace(/[^\\u4e00-\\u9fa5a-zA-Z0-9]/g, ' ').toLowerCase();\n  const tokens = [];\n  // 英文单词\n  const englishWords = cleaned.match(/[a-z0-9]+/g) || [];\n  tokens.push(...englishWords);\n  // 中文：bigram + trigram\n  const chinese = cleaned.replace(/[^\\u4e00-\\u9fa5]/g, '');\n  for (let i = 0; i < chinese.length - 1; i++) {\n    tokens.push(chinese.slice(i, i + 2));\n  }\n  for (let i = 0; i < chinese.length - 2; i++) {\n    tokens.push(chinese.slice(i, i + 3));\n  }\n  // 单字\n  for (const char of chinese) {\n    tokens.push(char);\n  }\n  // 同义词扩展\n  const expanded = [];\n  for (const token of tokens) {\n    expanded.push(token);\n    if (SYNONYMS[token]) {\n      expanded.push(...SYNONYMS[token]);\n    }\n  }\n  return expanded.filter(t => t.length > 0);\n}\n\n// BM25 索引\nclass BM25Index {\n  constructor() {\n    this.docs = []; // {id, tokens, length}\n    this.df = {};   // document frequency\n    this.totalDocs = 0;\n    this.avgDocLen = 0;\n  }\n\n  build(items) {\n    this.docs = [];\n    this.df = {};\n    this.totalDocs = items.length;\n    let totalLen = 0;\n\n    for (const item of items) {\n      const tokens = tokenize(item.title + ' ' + item.content);\n      this.docs.push({ id: item.id, tokens, length: tokens.length });\n      totalLen += tokens.length;\n\n      const uniqueTokens = new Set(tokens);\n      for (const t of uniqueTokens) {\n        this.df[t] = (this.df[t] || 0) + 1;\n      }\n    }\n    this.avgDocLen = this.docs.reduce((sum, d) => sum + d.length, 0) / this.totalDocs;\n  }\n\n  search(query, topK = 5) {\n    const queryTokens = tokenize(query);\n    if (queryTokens.length === 0) return [];\n\n    const scores = this.docs.map(doc => {\n      let score = 0;\n      const tf = {};\n      for (const t of doc.tokens) tf[t] = (tf[t] || 0) + 1;\n\n      for (const qt of queryTokens) {\n        if (!tf[qt]) continue;\n        const termFreq = tf[qt];\n        const docFreq = this.df[qt] || 0;\n        const idf = Math.log((this.totalDocs - docFreq + 0.5) / (docFreq + 0.5) + 1);\n        const tfNorm = (termFreq * (BM25_K1 + 1)) / (termFreq + BM25_K1 * (1 - BM25_B + BM25_B * doc.length / this.avgDocLen));\n        score += idf * tfNorm;\n      }\n      return { id: doc.id, score };\n    });\n\n    return scores.sort((a, b) => b.score - a.score).slice(0, topK);\n  }\n}\n\n// 向量索引（使用 sqlite-vec）\nclass VectorIndex {\n  constructor() {\n    this.dimension = 128; // 特征哈希维度\n  }\n\n  /**\n   * 简单特征哈希：将文本转为固定维度向量\n   * 用 SimHash 思想，对中文和英文都有效\n   */\n  textToVector(text) {\n    const tokens = tokenize(text);\n    const vector = new Float32Array(this.dimension);\n    for (let i = 0; i < tokens.length; i++) {\n      const token = tokens[i];\n      // 简单哈希\n      let hash = 0;\n      for (let j = 0; j < token.length; j++) {\n        hash = ((hash << 5) - hash + token.charCodeAt(j)) | 0;\n      }\n      const idx = Math.abs(hash) % this.dimension;\n      // 使用 +/-1 投影（特征哈希）\n      vector[idx] += (hash > 0 ? 1 : -1);\n    }\n    // L2 归一化\n    let norm = 0;\n    for (let i = 0; i < this.dimension; i++) norm += vector[i] * vector[i];\n    norm = Math.sqrt(norm) || 1;\n    for (let i = 0; i < this.dimension; i++) vector[i] /= norm;\n    return vector;\n  }\n\n  initTable(db) {\n    ensureVec(db);\n    db.exec(`\n      CREATE VIRTUAL TABLE IF NOT EXISTS knowledge_vec USING vec0(\n        id INTEGER PRIMARY KEY,\n        embedding float[${this.dimension}]\n      );\n    `);\n  }\n\n  upsert(db, id, text) {\n    const vector = this.textToVector(text);\n    const intId = Number(id);\n    if (isNaN(intId)) return;\n    // 先删除旧的\n    try { db.prepare('DELETE FROM knowledge_vec WHERE id = ?').run(intId); } catch {}\n    // 插入新的\n    db.prepare('INSERT INTO knowledge_vec (id, embedding) VALUES (?, ?)').run(intId, Buffer.from(vector.buffer));\n  }\n\n  search(db, queryText, topK = 5) {\n    ensureVec(db);\n    const queryVector = this.textToVector(queryText);\n    try {\n      const results = db.prepare(`\n        SELECT id, distance FROM knowledge_vec WHERE embedding MATCH ? ORDER BY distance LIMIT ?\n      `).all(Buffer.from(queryVector.buffer), topK);\n      return results.map(r => ({ id: r.id, score: 1 - r.distance })); // distance -> similarity\n    } catch (e) {\n      // 如果向量表还没建好，返回空\n      return [];\n    }\n  }\n}\n\n// 全局实例\nlet bm25Index = null;\nconst vectorIndex = new VectorIndex();\n\n/**\n * 重建索引\n */\nfunction rebuildIndex() {\n  const db = getDb();\n  const items = db.prepare(\"SELECT id, title, content FROM ai_knowledge WHERE status = 'active'\").all();\n\n  // 重建 BM25\n  bm25Index = new BM25Index();\n  bm25Index.build(items);\n\n  // 重建向量\n  try {\n    vectorIndex.initTable(db);\n    // 清空旧数据\n    try { db.prepare('DELETE FROM knowledge_vec').run(); } catch {}\n    for (const item of items) {\n      vectorIndex.upsert(db, item.id, item.title + ' ' + item.content);\n    }\n  } catch (e) {\n    console.warn('向量索引初始化失败（不影响 BM25 检索）:', e.message);\n  }\n\n  console.log(`RAG 索引已重建: ${items.length} 条知识`);\n  return items.length;\n}\n\n/**\n * 混合检索：BM25 + 向量，加权合并\n * @param {string} query - 用户问题\n * @param {number} topK - 返回条数\n * @returns {Array} [{id, title, content, score}]\n */\nfunction retrieve(query, topK = 5) {\n  const db = getDb();\n\n  // 确保索引已建\n  if (!bm25Index) rebuildIndex();\n\n  // BM25 检索\n  const bm25Results = bm25Index.search(query, topK * 2);\n\n  // 向量检索\n  let vecResults = [];\n  try {\n    vecResults = vectorIndex.search(db, query, topK * 2);\n  } catch {}\n\n  // 合并分数（加权：BM25 0.6 + 向量 0.4）\n  const scoreMap = {};\n  for (const r of bm25Results) {\n    scoreMap[r.id] = (scoreMap[r.id] || 0) + r.score * 0.6;\n  }\n  for (const r of vecResults) {\n    scoreMap[r.id] = (scoreMap[r.id] || 0) + r.score * 0.4;\n  }\n\n  // 排序取 Top-K\n  const sorted = Object.entries(scoreMap)\n    .map(([id, score]) => ({ id: Number(id), score }))\n    .sort((a, b) => b.score - a.score)\n    .slice(0, topK);\n\n  // 获取知识条目详情\n  if (sorted.length === 0) return [];\n  const placeholders = sorted.map(() => '?').join(',');\n  const items = db.prepare(`SELECT id, title, content, category FROM ai_knowledge WHERE id IN (${placeholders}) AND status = 'active'`)\n    .all(...sorted.map(s => s.id));\n\n  // 按分数排序\n  const itemMap = {};\n  for (const item of items) itemMap[item.id] = item;\n  return sorted\n    .filter(s => itemMap[s.id])\n    .map(s => ({ ...itemMap[s.id], score: s.score }));\n}\n\nmodule.exports = { retrieve, rebuildIndex, tokenize };\n"}}, "backend/services/ai.js": {"ff8d17b14023e5654fa002ec3900c26e454165c2": {"t": 1785746698.8499682, "n": 9416, "h": "634f819f92493b3e", "c": "/**\n * AI 客服服务 — 调用 LLM API 回答用户问题\n */\nconst https = require('https');\nconst { getDb } = require('../database/schema');\nconst { retrieve: ragRetrieve, rebuildIndex } = require('./rag');\n\n/**\n * 从 settings 获取 AI 配置\n */\nfunction getAIConfig() {\n  const db = getDb();\n  const rows = db.prepare(\"SELECT key, value FROM settings WHERE key IN ('ai_provider','ai_api_key','ai_model','ai_system_prompt','ai_base_url')\").all();\n  const config = {};\n  for (const row of rows) config[row.key] = row.value;\n  return config;\n}\n\n/**\n * RAG 检索：根据用户问题检索相关知识（支持图片）\n */\nasync function getKnowledgeContext(query) {\n  if (!query) return '';\n  try {\n    const results = await ragRetrieve(query, 5);\n    if (results.length === 0) return '';\n    let context = '以下是与用户问题最相关的知识库内容，请基于这些信息回答：\\n\\n';\n    for (const item of results) {\n      context += `【${item.category || '未分类'}】${item.title}\\n${item.content}\\n`;\n      // 附带来源链接（教程/FAQ）\n      if (item.source === 'tutorial' && item.parentId) {\n        context += `📖 查看教程详情：[点击这里](/tutorials/${item.parentId})\\n`;\n      } else if (item.source === 'faq') {\n        context += `❓ 更多常见问题：[查看FAQ](/faq)\\n`;\n      }\n      // 提取内容中的外部链接\n      const urls = (item.content || '').match(/https?:\\/\\/[^\\s\"'<>)\\]]+/g) || [];\n      const uniqueUrls = [...new Set(urls)];\n      if (uniqueUrls.length > 0) {\n        context += `🔗 相关链接：${uniqueUrls.map(u => `[${u}](${u})`).join('、')}\\n`;\n      }\n      // 附带相关图片\n      if (item.images && item.images.length > 0) {\n        context += `相关图片：${item.images.join(', ')}\\n`;\n      }\n      context += '\\n';\n    }\n    return context;\n  } catch (e) {\n    console.warn('RAG 检索失败:', e.message);\n    return '';\n  }\n}\n\n/**\n * 调用 OpenAI 兼容 API\n */\nfunction callOpenAICompat(baseUrl, apiKey, model, messages) {\n  return new Promise((resolve, reject) => {\n    const parsed = new URL(baseUrl);\n    const body = JSON.stringify({\n      model,\n      messages,\n      temperature: 0.7,\n      max_tokens: 2000,\n    });\n    const options = {\n      hostname: parsed.hostname,\n      port: parsed.port || 443,\n      path: parsed.pathname.replace(/\\/$/, '') + '/chat/completions',\n      method: 'POST',\n      headers: {\n        'Content-Type': 'application/json',\n        'Authorization': `Bearer ${apiKey}`,\n        'Content-Length': Buffer.byteLength(body),\n      },\n    };\n    const req = https.request(options, res => {\n      let data = '';\n      res.on('data', c => data += c);\n      res.on('end', () => {\n        try {\n          const json = JSON.parse(data);\n          if (json.error) return reject(new Error(json.error.message || JSON.stringify(json.error)));\n          const content = json.choices?.[0]?.message?.content || '';\n          resolve(content);\n        } catch (e) {\n          reject(new Error('AI 响应解析失败: ' + data.slice(0, 200)));\n        }\n      });\n    });\n    req.on('error', reject);\n    req.write(body);\n    req.end();\n  });\n}\n\n/**\n * 获取 provider 的 baseUrl\n */\nfunction getProviderBaseUrl(provider) {\n  const urls = {\n    deepseek: 'https://api.deepseek.com',\n    openai: 'https://api.openai.com/v1',\n    qwen: 'https://dashscope.aliyuncs.com/compatible-mode/v1',\n  };\n  return urls[provider] || '';\n}\n\n/**\n * AI 对话主函数（三层记忆：用户画像 + 历史对话 + 全局知识库）\n */\nasync function chat(history, userMessage, imageUrl = '', options = {}) {\n  const config = getAIConfig();\n  if (!config.ai_provider || !config.ai_api_key) throw new Error('AI 客服未配置，请在后台设置中配置 AI 服务商和 API Key');\n  const baseUrl = config.ai_base_url || getProviderBaseUrl(config.ai_provider);\n  if (!baseUrl) throw new Error('未知的 AI 服务商: ' + config.ai_provider);\n  const model = config.ai_model || 'deepseek-chat';\n\n  let systemPrompt = config.ai_system_prompt || '你是 imai.work 的智能客服助手，专门解答关于养号、获客、短视频运营的问题。请用中文回答，语气友好专业。如果不确定答案，请诚实说明并建议用户联系人工客服。';\n\n  // 记忆 1：用户画像（静态+动态）\n  systemPrompt += getUserProfile(options.userId);\n\n  // 记忆 2：历史对话记忆\n  systemPrompt += getConversationMemory(options.userId);\n\n  // 记忆 3：RAG 知识库检索\n  const knowledge = await getKnowledgeContext(userMessage);\n  if (knowledge) systemPrompt += '\\n\\n' + knowledge;\n\n  systemPrompt += '\\n\\n【回复规则】\\n1. 如果知识库中有相关图片，请在回答中用 markdown 图片语法展示。\\n2. 如果知识库内容中带有 📖 教程链接、🔗 外部链接 或 ❓ FAQ链接，你必须在回答末尾原样附上这些链接，不要修改链接地址。\\n3. 回答要简洁实用，避免过多废话。';\n\n  const messages = [{ role: 'system', content: systemPrompt }];\n  const recentHistory = history.slice(-20);\n  for (const msg of recentHistory) messages.push({ role: msg.role, content: msg.content });\n\n  if (imageUrl) {\n    messages.push({ role: 'user', content: [{ type: 'text', text: userMessage || '请分析这张图片' }, { type: 'image_url', image_url: { url: imageUrl } }] });\n  } else {\n    messages.push({ role: 'user', content: userMessage });\n  }\n\n  return callOpenAICompat(baseUrl, config.ai_api_key, model, messages);\n}\n\n/**\n * 获取历史对话记忆（最近 5 条对话的摘要）\n */\nfunction getConversationMemory(userId) {\n  if (!userId) return '';\n  try {\n    const db = getDb();\n    const convs = db.prepare(\n      \"SELECT summary FROM ai_conversations WHERE user_id = ? AND status = 'closed' AND summary != '' ORDER BY updated_at DESC LIMIT 5\"\n    ).all(userId);\n    if (convs.length === 0) return '';\n    let memory = '\\n\\n【历史记忆】用户之前的对话摘要：\\n';\n    for (const c of convs) memory += '- ' + c.summary + '\\n';\n    memory += '请参考这些历史信息，避免重复回答已解决的问题。';\n    return memory;\n  } catch { return ''; }\n}\n\n/**\n * 获取用户画像记忆（静态 + 动态）\n */\nfunction getUserProfile(userId) {\n  if (!userId) return '';\n  try {\n    const db = getDb();\n    const parts = [];\n\n    // 静态信息：注册时填写的\n    const user = db.prepare('SELECT nickname, company, company_role, industry FROM users WHERE id = ?').get(userId);\n    if (user) {\n      if (user.company) parts.push('公司：' + user.company);\n      if (user.company_role) parts.push('职位：' + user.company_role);\n      if (user.industry) parts.push('行业：' + user.industry);\n    }\n\n    // 动态记忆：从对话中学习的\n    const memories = db.prepare(\n      \"SELECT category, content FROM user_memory WHERE user_id = ? ORDER BY confidence DESC, updated_at DESC LIMIT 15\"\n    ).all(userId);\n    if (memories.length > 0) {\n      const catNames = { interest: '兴趣需求', behavior: '行为特征', context: '业务背景', preference: '偏好习惯', general: '其他' };\n      const byCategory = {};\n      for (const m of memories) {\n        const cat = catNames[m.category] || '其他';\n        if (!byCategory[cat]) byCategory[cat] = [];\n        byCategory[cat].push(m.content);\n      }\n      for (const [cat, items] of Object.entries(byCategory)) {\n        parts.push(cat + '：' + items.join('、'));\n      }\n    }\n\n    if (parts.length === 0) return '';\n    return '\\n\\n【用户画像】' + parts.join('；') + '。请根据用户的具体情况给出针对性回答，不要复述这些信息。';\n  } catch { return ''; }\n}\n\n/**\n * 从对话中提取用户特征并保存\n */\nasync function extractUserMemory(userId, conversationId, history) {\n  if (!userId || history.length < 2) return;\n  const config = getAIConfig();\n  if (!config.ai_provider || !config.ai_api_key) return;\n  const baseUrl = config.ai_base_url || getProviderBaseUrl(config.ai_provider);\n  if (!baseUrl) return;\n  const model = config.ai_model || 'deepseek-chat';\n\n  const conversationText = history.map(m => (m.role === 'user' ? '用户' : 'AI') + '：' + m.content).join('\\n');\n\n  try {\n    const result = await callOpenAICompat(baseUrl, config.ai_api_key, model, [\n      { role: 'system', content: '你是一个用户画像分析器。从对话中提取用户的特征信息，输出 JSON 数组。\\n\\n每个条目格式：{\"category\": \"分类\", \"content\": \"特征描述\"}\\n分类包括：\\n- interest: 用户的兴趣需求（如\"对抖音养号感兴趣\"）\\n- behavior: 用户的行为特征（如\"有多个账号需要管理\"）\\n- context: 用户的业务背景（如\"做餐饮行业\"、\"是个体经营者\"）\\n- preference: 用户的偏好习惯（如\"喜欢简洁的回答\"）\\n\\n规则：\\n1. 只提取明确的信息，不要猜测\\n2. 每条不超过30字\\n3. 最多提取5条\\n4. 如果没有值得记录的特征，输出空数组 []\\n5. 只输出 JSON，不要其他内容' },\n      { role: 'user', content: conversationText.slice(-3000) },\n    ]);\n\n    const jsonMatch = result.match(/\\[.*\\]/s);\n    if (!jsonMatch) return;\n    const traits = JSON.parse(jsonMatch[0]);\n    if (!Array.isArray(traits) || traits.length === 0) return;\n\n    const db = getDb();\n    const upsert = db.prepare(\n      \"INSERT INTO user_memory (user_id, category, content, source_conversation_id) VALUES (?, ?, ?, ?) ON CONFLICT(user_id, content) DO UPDATE SET confidence = confidence + 0.2, updated_at = datetime('now','localtime')\"\n    );\n\n    let saved = 0;\n    for (const trait of traits) {\n      if (!trait.content || trait.content.length > 50) continue;\n      const validCategories = ['interest', 'behavior', 'context', 'preference', 'general'];\n      const cat = validCategories.includes(trait.category) ? trait.category : 'general';\n      try { upsert.run(userId, cat, trait.content, conversationId); saved++; } catch {}\n    }\n    if (saved > 0) console.log('🧠 用户画像更新: 用户' + userId + ' 新增' + saved + '条特征');\n  } catch (e) {\n    console.warn('用户画像提取失败:', e.message);\n  }\n}\n\n/**\n * 生成对话摘要\n */\nasync function generateSummary(history) {\n  const config = getAIConfig();\n  if (!config.ai_provider || !config.ai_api_key) return '';\n  const baseUrl = config.ai_base_url || getProviderBaseUrl(config.ai_provider);\n  if (!baseUrl) return '';\n  const model = config.ai_model || 'deepseek-chat';\n\n  const conversationText = history.map(m => (m.role === 'user' ? '用户' : 'AI') + '：' + m.content).join('\\n');\n\n  try {\n    const summary = await callOpenAICompat(baseUrl, config.ai_api_key, model, [\n      { role: 'system', content: '请用一句话（不超过50字）概括这段对话的核心内容，用于记忆存档。只输出摘要，不要其他内容。' },\n      { role: 'user', content: conversationText.slice(-2000) },\n    ]);\n    return summary.trim().slice(0, 200);\n  } catch {\n    return '';\n  }\n}\n\nmodule.exports = { chat, getAIConfig, generateSummary, extractUserMemory };\n"}, "54c84ea54c3a90233572722cf5e1305947594777": {"t": 1785766724.3817883, "n": 8823, "h": "1afc4fd76b673d90", "c": "/**\n * AI 客服服务 — 调用 LLM API 回答用户问题\n */\nconst https = require('https');\nconst { getDb } = require('../database/schema');\nconst { retrieve: ragRetrieve, rebuildIndex } = require('./rag');\n\n/**\n * 从 settings 获取 AI 配置\n */\nfunction getAIConfig() {\n  const db = getDb();\n  const rows = db.prepare(\"SELECT key, value FROM settings WHERE key IN ('ai_provider','ai_api_key','ai_model','ai_system_prompt','ai_base_url')\").all();\n  const config = {};\n  for (const row of rows) config[row.key] = row.value;\n  return config;\n}\n\n/**\n * RAG 检索：根据用户问题检索相关知识（支持图片）\n */\nasync function getKnowledgeContext(query) {\n  if (!query) return '';\n  try {\n    const results = await ragRetrieve(query, 5);\n    if (results.length === 0) return '';\n    let context = '以下是与用户问题最相关的知识库内容，请基于这些信息回答：\\n\\n';\n    for (const item of results) {\n      context += `【${item.category || '未分类'}】${item.title}\\n${item.content}\\n`;\n      // 附带相关图片\n      if (item.images && item.images.length > 0) {\n        context += `相关图片：${item.images.join(', ')}\\n`;\n      }\n      context += '\\n';\n    }\n    return context;\n  } catch (e) {\n    console.warn('RAG 检索失败:', e.message);\n    return '';\n  }\n}\n\n/**\n * 调用 OpenAI 兼容 API\n */\nfunction callOpenAICompat(baseUrl, apiKey, model, messages) {\n  return new Promise((resolve, reject) => {\n    const parsed = new URL(baseUrl);\n    const body = JSON.stringify({\n      model,\n      messages,\n      temperature: 0.7,\n      max_tokens: 2000,\n    });\n    const options = {\n      hostname: parsed.hostname,\n      port: parsed.port || 443,\n      path: parsed.pathname.replace(/\\/$/, '') + '/chat/completions',\n      method: 'POST',\n      headers: {\n        'Content-Type': 'application/json',\n        'Authorization': `Bearer ${apiKey}`,\n        'Content-Length': Buffer.byteLength(body),\n      },\n    };\n    const req = https.request(options, res => {\n      let data = '';\n      res.on('data', c => data += c);\n      res.on('end', () => {\n        try {\n          const json = JSON.parse(data);\n          if (json.error) return reject(new Error(json.error.message || JSON.stringify(json.error)));\n          const content = json.choices?.[0]?.message?.content || '';\n          resolve(content);\n        } catch (e) {\n          reject(new Error('AI 响应解析失败: ' + data.slice(0, 200)));\n        }\n      });\n    });\n    req.on('error', reject);\n    req.write(body);\n    req.end();\n  });\n}\n\n/**\n * 获取 provider 的 baseUrl\n */\nfunction getProviderBaseUrl(provider) {\n  const urls = {\n    deepseek: 'https://api.deepseek.com',\n    openai: 'https://api.openai.com/v1',\n    qwen: 'https://dashscope.aliyuncs.com/compatible-mode/v1',\n  };\n  return urls[provider] || '';\n}\n\n/**\n * AI 对话主函数（三层记忆：用户画像 + 历史对话 + 全局知识库）\n */\nasync function chat(history, userMessage, imageUrl = '', options = {}) {\n  const config = getAIConfig();\n  if (!config.ai_provider || !config.ai_api_key) throw new Error('AI 客服未配置，请在后台设置中配置 AI 服务商和 API Key');\n  const baseUrl = config.ai_base_url || getProviderBaseUrl(config.ai_provider);\n  if (!baseUrl) throw new Error('未知的 AI 服务商: ' + config.ai_provider);\n  const model = config.ai_model || 'deepseek-chat';\n\n  let systemPrompt = config.ai_system_prompt || '你是 imai.work 的智能客服助手，专门解答关于养号、获客、短视频运营的问题。请用中文回答，语气友好专业。如果不确定答案，请诚实说明并建议用户联系人工客服。';\n\n  // 记忆 1：用户画像（静态+动态）\n  systemPrompt += getUserProfile(options.userId);\n\n  // 记忆 2：历史对话记忆\n  systemPrompt += getConversationMemory(options.userId);\n\n  // 记忆 3：RAG 知识库检索\n  const knowledge = await getKnowledgeContext(userMessage);\n  if (knowledge) systemPrompt += '\\n\\n' + knowledge;\n\n  systemPrompt += '\\n\\n【回复规则】如果知识库中有相关图片，请在回答中用 markdown 图片语法展示。回答要简洁实用，避免过多废话。';\n\n  const messages = [{ role: 'system', content: systemPrompt }];\n  const recentHistory = history.slice(-20);\n  for (const msg of recentHistory) messages.push({ role: msg.role, content: msg.content });\n\n  if (imageUrl) {\n    messages.push({ role: 'user', content: [{ type: 'text', text: userMessage || '请分析这张图片' }, { type: 'image_url', image_url: { url: imageUrl } }] });\n  } else {\n    messages.push({ role: 'user', content: userMessage });\n  }\n\n  return callOpenAICompat(baseUrl, config.ai_api_key, model, messages);\n}\n\n/**\n * 获取历史对话记忆（最近 5 条对话的摘要）\n */\nfunction getConversationMemory(userId) {\n  if (!userId) return '';\n  try {\n    const db = getDb();\n    const convs = db.prepare(\n      \"SELECT summary FROM ai_conversations WHERE user_id = ? AND status = 'closed' AND summary != '' ORDER BY updated_at DESC LIMIT 5\"\n    ).all(userId);\n    if (convs.length === 0) return '';\n    let memory = '\\n\\n【历史记忆】用户之前的对话摘要：\\n';\n    for (const c of convs) memory += '- ' + c.summary + '\\n';\n    memory += '请参考这些历史信息，避免重复回答已解决的问题。';\n    return memory;\n  } catch { return ''; }\n}\n\n/**\n * 获取用户画像记忆（静态 + 动态）\n */\nfunction getUserProfile(userId) {\n  if (!userId) return '';\n  try {\n    const db = getDb();\n    const parts = [];\n\n    // 静态信息：注册时填写的\n    const user = db.prepare('SELECT nickname, company, company_role, industry FROM users WHERE id = ?').get(userId);\n    if (user) {\n      if (user.company) parts.push('公司：' + user.company);\n      if (user.company_role) parts.push('职位：' + user.company_role);\n      if (user.industry) parts.push('行业：' + user.industry);\n    }\n\n    // 动态记忆：从对话中学习的\n    const memories = db.prepare(\n      \"SELECT category, content FROM user_memory WHERE user_id = ? ORDER BY confidence DESC, updated_at DESC LIMIT 15\"\n    ).all(userId);\n    if (memories.length > 0) {\n      const catNames = { interest: '兴趣需求', behavior: '行为特征', context: '业务背景', preference: '偏好习惯', general: '其他' };\n      const byCategory = {};\n      for (const m of memories) {\n        const cat = catNames[m.category] || '其他';\n        if (!byCategory[cat]) byCategory[cat] = [];\n        byCategory[cat].push(m.content);\n      }\n      for (const [cat, items] of Object.entries(byCategory)) {\n        parts.push(cat + '：' + items.join('、'));\n      }\n    }\n\n    if (parts.length === 0) return '';\n    return '\\n\\n【用户画像】' + parts.join('；') + '。请根据用户的具体情况给出针对性回答，不要复述这些信息。';\n  } catch { return ''; }\n}\n\n/**\n * 从对话中提取用户特征并保存\n */\nasync function extractUserMemory(userId, conversationId, history) {\n  if (!userId || history.length < 2) return;\n  const config = getAIConfig();\n  if (!config.ai_provider || !config.ai_api_key) return;\n  const baseUrl = config.ai_base_url || getProviderBaseUrl(config.ai_provider);\n  if (!baseUrl) return;\n  const model = config.ai_model || 'deepseek-chat';\n\n  const conversationText = history.map(m => (m.role === 'user' ? '用户' : 'AI') + '：' + m.content).join('\\n');\n\n  try {\n    const result = await callOpenAICompat(baseUrl, config.ai_api_key, model, [\n      { role: 'system', content: '你是一个用户画像分析器。从对话中提取用户的特征信息，输出 JSON 数组。\\n\\n每个条目格式：{\"category\": \"分类\", \"content\": \"特征描述\"}\\n分类包括：\\n- interest: 用户的兴趣需求（如\"对抖音养号感兴趣\"）\\n- behavior: 用户的行为特征（如\"有多个账号需要管理\"）\\n- context: 用户的业务背景（如\"做餐饮行业\"、\"是个体经营者\"）\\n- preference: 用户的偏好习惯（如\"喜欢简洁的回答\"）\\n\\n规则：\\n1. 只提取明确的信息，不要猜测\\n2. 每条不超过30字\\n3. 最多提取5条\\n4. 如果没有值得记录的特征，输出空数组 []\\n5. 只输出 JSON，不要其他内容' },\n      { role: 'user', content: conversationText.slice(-3000) },\n    ]);\n\n    const jsonMatch = result.match(/\\[.*\\]/s);\n    if (!jsonMatch) return;\n    const traits = JSON.parse(jsonMatch[0]);\n    if (!Array.isArray(traits) || traits.length === 0) return;\n\n    const db = getDb();\n    const upsert = db.prepare(\n      \"INSERT INTO user_memory (user_id, category, content, source_conversation_id) VALUES (?, ?, ?, ?) ON CONFLICT(user_id, content) DO UPDATE SET confidence = confidence + 0.2, updated_at = datetime('now','localtime')\"\n    );\n\n    let saved = 0;\n    for (const trait of traits) {\n      if (!trait.content || trait.content.length > 50) continue;\n      const validCategories = ['interest', 'behavior', 'context', 'preference', 'general'];\n      const cat = validCategories.includes(trait.category) ? trait.category : 'general';\n      try { upsert.run(userId, cat, trait.content, conversationId); saved++; } catch {}\n    }\n    if (saved > 0) console.log('🧠 用户画像更新: 用户' + userId + ' 新增' + saved + '条特征');\n  } catch (e) {\n    console.warn('用户画像提取失败:', e.message);\n  }\n}\n\n/**\n * 生成对话摘要\n */\nasync function generateSummary(history) {\n  const config = getAIConfig();\n  if (!config.ai_provider || !config.ai_api_key) return '';\n  const baseUrl = config.ai_base_url || getProviderBaseUrl(config.ai_provider);\n  if (!baseUrl) return '';\n  const model = config.ai_model || 'deepseek-chat';\n\n  const conversationText = history.map(m => (m.role === 'user' ? '用户' : 'AI') + '：' + m.content).join('\\n');\n\n  try {\n    const summary = await callOpenAICompat(baseUrl, config.ai_api_key, model, [\n      { role: 'system', content: '请用一句话（不超过50字）概括这段对话的核心内容，用于记忆存档。只输出摘要，不要其他内容。' },\n      { role: 'user', content: conversationText.slice(-2000) },\n    ]);\n    return summary.trim().slice(0, 200);\n  } catch {\n    return '';\n  }\n}\n\nmodule.exports = { chat, getAIConfig, generateSummary, extractUserMemory };\n"}, "07d9a4cb735c78690d56faa55797451246cb6559": {"t": 1785766725.6166017, "n": 4338, "h": "51d6b60f42418b46", "c": "/**\n * AI 客服服务 — 调用 LLM API 回答用户问题\n */\nconst https = require('https');\nconst { getDb } = require('../database/schema');\nconst { retrieve: ragRetrieve, rebuildIndex } = require('./rag');\n\n/**\n * 从 settings 获取 AI 配置\n */\nfunction getAIConfig() {\n  const db = getDb();\n  const rows = db.prepare(\"SELECT key, value FROM settings WHERE key IN ('ai_provider','ai_api_key','ai_model','ai_system_prompt','ai_base_url')\").all();\n  const config = {};\n  for (const row of rows) config[row.key] = row.value;\n  return config;\n}\n\n/**\n * RAG 检索：根据用户问题检索相关知识（支持图片）\n */\nasync function getKnowledgeContext(query) {\n  if (!query) return '';\n  try {\n    const results = await ragRetrieve(query, 5);\n    if (results.length === 0) return '';\n    let context = '以下是与用户问题最相关的知识库内容，请基于这些信息回答：\\n\\n';\n    for (const item of results) {\n      context += `【${item.category || '未分类'}】${item.title}\\n${item.content}\\n`;\n      // 附带相关图片\n      if (item.images && item.images.length > 0) {\n        context += `相关图片：${item.images.join(', ')}\\n`;\n      }\n      context += '\\n';\n    }\n    return context;\n  } catch (e) {\n    console.warn('RAG 检索失败:', e.message);\n    return '';\n  }\n}\n\n/**\n * 调用 OpenAI 兼容 API\n */\nfunction callOpenAICompat(baseUrl, apiKey, model, messages) {\n  return new Promise((resolve, reject) => {\n    const parsed = new URL(baseUrl);\n    const body = JSON.stringify({\n      model,\n      messages,\n      temperature: 0.7,\n      max_tokens: 2000,\n    });\n    const options = {\n      hostname: parsed.hostname,\n      port: parsed.port || 443,\n      path: parsed.pathname.replace(/\\/$/, '') + '/chat/completions',\n      method: 'POST',\n      headers: {\n        'Content-Type': 'application/json',\n        'Authorization': `Bearer ${apiKey}`,\n        'Content-Length': Buffer.byteLength(body),\n      },\n    };\n    const req = https.request(options, res => {\n      let data = '';\n      res.on('data', c => data += c);\n      res.on('end', () => {\n        try {\n          const json = JSON.parse(data);\n          if (json.error) return reject(new Error(json.error.message || JSON.stringify(json.error)));\n          const content = json.choices?.[0]?.message?.content || '';\n          resolve(content);\n        } catch (e) {\n          reject(new Error('AI 响应解析失败: ' + data.slice(0, 200)));\n        }\n      });\n    });\n    req.on('error', reject);\n    req.write(body);\n    req.end();\n  });\n}\n\n/**\n * 获取 provider 的 baseUrl\n */\nfunction getProviderBaseUrl(provider) {\n  const urls = {\n    deepseek: 'https://api.deepseek.com',\n    openai: 'https://api.openai.com/v1',\n    qwen: 'https://dashscope.aliyuncs.com/compatible-mode/v1',\n  };\n  return urls[provider] || '';\n}\n\n/**\n * AI 对话主函数\n * @param {Array} history - 对话历史 [{role, content}]\n * @param {string} userMessage - 用户新消息\n * @param {string} imageUrl - 用户上传的图片 URL（可选）\n * @returns {string} AI 回复\n */\nasync function chat(history, userMessage, imageUrl = '') {\n  const config = getAIConfig();\n\n  if (!config.ai_provider || !config.ai_api_key) {\n    throw new Error('AI 客服未配置，请在后台设置中配置 AI 服务商和 API Key');\n  }\n\n  const baseUrl = config.ai_base_url || getProviderBaseUrl(config.ai_provider);\n  if (!baseUrl) throw new Error('未知的 AI 服务商: ' + config.ai_provider);\n\n  const model = config.ai_model || 'deepseek-chat';\n\n  // 构建 system prompt\n  let systemPrompt = config.ai_system_prompt || '你是 imai.work 的智能客服助手，专门解答关于养号、获客、短视频运营的问题。请用中文回答，语气友好专业。如果不确定答案，请诚实说明并建议用户联系人工客服。';\n\n  // RAG 检索相关知识\n  const knowledge = await getKnowledgeContext(userMessage);\n  if (knowledge) {\n    systemPrompt += '\\n\\n' + knowledge;\n  }\n\n  // 告诉 LLM 可以输出图片\n  systemPrompt += '\\n\\n【回复规则】如果知识库中有相关图片，请在回答中用 markdown 图片语法展示：![描述](图片URL)。图片能让用户更直观地理解。';\n\n  // 构建消息列表\n  const messages = [{ role: 'system', content: systemPrompt }];\n\n  // 加入对话历史（最近 10 轮）\n  const recentHistory = history.slice(-20);\n  for (const msg of recentHistory) {\n    messages.push({ role: msg.role, content: msg.content });\n  }\n\n  // 加入用户新消息（支持图片）\n  if (imageUrl) {\n    messages.push({\n      role: 'user',\n      content: [\n        { type: 'text', text: userMessage || '请分析这张图片' },\n        { type: 'image_url', image_url: { url: imageUrl } },\n      ],\n    });\n  } else {\n    messages.push({ role: 'user', content: userMessage });\n  }\n\n  const reply = await callOpenAICompat(baseUrl, config.ai_api_key, model, messages);\n  return reply;\n}\n\nmodule.exports = { chat, getAIConfig };\n"}, "3434aeeecff5dee82ac201362a02222b99c273e2": {"t": 1785766727.0413268, "n": 4320, "h": "f6a7fee5e56789f7", "c": "/**\n * AI 客服服务 — 调用 LLM API 回答用户问题\n */\nconst https = require('https');\nconst { getDb } = require('../database/schema');\nconst { retrieve: ragRetrieve, rebuildIndex } = require('./rag');\n\n/**\n * 从 settings 获取 AI 配置\n */\nfunction getAIConfig() {\n  const db = getDb();\n  const rows = db.prepare(\"SELECT key, value FROM settings WHERE key IN ('ai_provider','ai_api_key','ai_model','ai_system_prompt','ai_base_url')\").all();\n  const config = {};\n  for (const row of rows) config[row.key] = row.value;\n  return config;\n}\n\n/**\n * RAG 检索：根据用户问题检索相关知识（支持图片）\n */\nfunction getKnowledgeContext(query) {\n  if (!query) return '';\n  try {\n    const results = ragRetrieve(query, 5);\n    if (results.length === 0) return '';\n    let context = '以下是与用户问题最相关的知识库内容，请基于这些信息回答：\\n\\n';\n    for (const item of results) {\n      context += `【${item.category || '未分类'}】${item.title}\\n${item.content}\\n`;\n      // 附带相关图片\n      if (item.images && item.images.length > 0) {\n        context += `相关图片：${item.images.join(', ')}\\n`;\n      }\n      context += '\\n';\n    }\n    return context;\n  } catch (e) {\n    console.warn('RAG 检索失败:', e.message);\n    return '';\n  }\n}\n\n/**\n * 调用 OpenAI 兼容 API\n */\nfunction callOpenAICompat(baseUrl, apiKey, model, messages) {\n  return new Promise((resolve, reject) => {\n    const parsed = new URL(baseUrl);\n    const body = JSON.stringify({\n      model,\n      messages,\n      temperature: 0.7,\n      max_tokens: 2000,\n    });\n    const options = {\n      hostname: parsed.hostname,\n      port: parsed.port || 443,\n      path: parsed.pathname.replace(/\\/$/, '') + '/chat/completions',\n      method: 'POST',\n      headers: {\n        'Content-Type': 'application/json',\n        'Authorization': `Bearer ${apiKey}`,\n        'Content-Length': Buffer.byteLength(body),\n      },\n    };\n    const req = https.request(options, res => {\n      let data = '';\n      res.on('data', c => data += c);\n      res.on('end', () => {\n        try {\n          const json = JSON.parse(data);\n          if (json.error) return reject(new Error(json.error.message || JSON.stringify(json.error)));\n          const content = json.choices?.[0]?.message?.content || '';\n          resolve(content);\n        } catch (e) {\n          reject(new Error('AI 响应解析失败: ' + data.slice(0, 200)));\n        }\n      });\n    });\n    req.on('error', reject);\n    req.write(body);\n    req.end();\n  });\n}\n\n/**\n * 获取 provider 的 baseUrl\n */\nfunction getProviderBaseUrl(provider) {\n  const urls = {\n    deepseek: 'https://api.deepseek.com',\n    openai: 'https://api.openai.com/v1',\n    qwen: 'https://dashscope.aliyuncs.com/compatible-mode/v1',\n  };\n  return urls[provider] || '';\n}\n\n/**\n * AI 对话主函数\n * @param {Array} history - 对话历史 [{role, content}]\n * @param {string} userMessage - 用户新消息\n * @param {string} imageUrl - 用户上传的图片 URL（可选）\n * @returns {string} AI 回复\n */\nasync function chat(history, userMessage, imageUrl = '') {\n  const config = getAIConfig();\n\n  if (!config.ai_provider || !config.ai_api_key) {\n    throw new Error('AI 客服未配置，请在后台设置中配置 AI 服务商和 API Key');\n  }\n\n  const baseUrl = config.ai_base_url || getProviderBaseUrl(config.ai_provider);\n  if (!baseUrl) throw new Error('未知的 AI 服务商: ' + config.ai_provider);\n\n  const model = config.ai_model || 'deepseek-chat';\n\n  // 构建 system prompt\n  let systemPrompt = config.ai_system_prompt || '你是 imai.work 的智能客服助手，专门解答关于养号、获客、短视频运营的问题。请用中文回答，语气友好专业。如果不确定答案，请诚实说明并建议用户联系人工客服。';\n\n  // RAG 检索相关知识\n  const knowledge = getKnowledgeContext(userMessage);\n  if (knowledge) {\n    systemPrompt += '\\n\\n' + knowledge;\n  }\n\n  // 告诉 LLM 可以输出图片\n  systemPrompt += '\\n\\n【回复规则】如果知识库中有相关图片，请在回答中用 markdown 图片语法展示：![描述](图片URL)。图片能让用户更直观地理解。';\n\n  // 构建消息列表\n  const messages = [{ role: 'system', content: systemPrompt }];\n\n  // 加入对话历史（最近 10 轮）\n  const recentHistory = history.slice(-20);\n  for (const msg of recentHistory) {\n    messages.push({ role: msg.role, content: msg.content });\n  }\n\n  // 加入用户新消息（支持图片）\n  if (imageUrl) {\n    messages.push({\n      role: 'user',\n      content: [\n        { type: 'text', text: userMessage || '请分析这张图片' },\n        { type: 'image_url', image_url: { url: imageUrl } },\n      ],\n    });\n  } else {\n    messages.push({ role: 'user', content: userMessage });\n  }\n\n  const reply = await callOpenAICompat(baseUrl, config.ai_api_key, model, messages);\n  return reply;\n}\n\nmodule.exports = { chat, getAIConfig };\n"}, "f882ddb355bf9ea7f8eff6c1d5cd683e145d0aec": {"t": 1785766728.2832396, "n": 4050, "h": "b560c8f976527d4e", "c": "/**\n * AI 客服服务 — 调用 LLM API 回答用户问题\n */\nconst https = require('https');\nconst { getDb } = require('../database/schema');\nconst { retrieve: ragRetrieve, rebuildIndex } = require('./rag');\n\n/**\n * 从 settings 获取 AI 配置\n */\nfunction getAIConfig() {\n  const db = getDb();\n  const rows = db.prepare(\"SELECT key, value FROM settings WHERE key IN ('ai_provider','ai_api_key','ai_model','ai_system_prompt','ai_base_url')\").all();\n  const config = {};\n  for (const row of rows) config[row.key] = row.value;\n  return config;\n}\n\n/**\n * RAG 检索：根据用户问题检索相关知识\n */\nfunction getKnowledgeContext(query) {\n  if (!query) return '';\n  try {\n    const results = ragRetrieve(query, 5);\n    if (results.length === 0) return '';\n    let context = '以下是与用户问题最相关的知识库内容，请基于这些信息回答：\\n\\n';\n    for (const item of results) {\n      context += `【${item.category || '未分类'}】${item.title}\\n${item.content}\\n\\n`;\n    }\n    return context;\n  } catch (e) {\n    console.warn('RAG 检索失败:', e.message);\n    return '';\n  }\n}\n\n/**\n * 调用 OpenAI 兼容 API\n */\nfunction callOpenAICompat(baseUrl, apiKey, model, messages) {\n  return new Promise((resolve, reject) => {\n    const parsed = new URL(baseUrl);\n    const body = JSON.stringify({\n      model,\n      messages,\n      temperature: 0.7,\n      max_tokens: 2000,\n    });\n    const options = {\n      hostname: parsed.hostname,\n      port: parsed.port || 443,\n      path: parsed.pathname.replace(/\\/$/, '') + '/chat/completions',\n      method: 'POST',\n      headers: {\n        'Content-Type': 'application/json',\n        'Authorization': `Bearer ${apiKey}`,\n        'Content-Length': Buffer.byteLength(body),\n      },\n    };\n    const req = https.request(options, res => {\n      let data = '';\n      res.on('data', c => data += c);\n      res.on('end', () => {\n        try {\n          const json = JSON.parse(data);\n          if (json.error) return reject(new Error(json.error.message || JSON.stringify(json.error)));\n          const content = json.choices?.[0]?.message?.content || '';\n          resolve(content);\n        } catch (e) {\n          reject(new Error('AI 响应解析失败: ' + data.slice(0, 200)));\n        }\n      });\n    });\n    req.on('error', reject);\n    req.write(body);\n    req.end();\n  });\n}\n\n/**\n * 获取 provider 的 baseUrl\n */\nfunction getProviderBaseUrl(provider) {\n  const urls = {\n    deepseek: 'https://api.deepseek.com',\n    openai: 'https://api.openai.com/v1',\n    qwen: 'https://dashscope.aliyuncs.com/compatible-mode/v1',\n  };\n  return urls[provider] || '';\n}\n\n/**\n * AI 对话主函数\n * @param {Array} history - 对话历史 [{role, content}]\n * @param {string} userMessage - 用户新消息\n * @param {string} imageUrl - 用户上传的图片 URL（可选）\n * @returns {string} AI 回复\n */\nasync function chat(history, userMessage, imageUrl = '') {\n  const config = getAIConfig();\n\n  if (!config.ai_provider || !config.ai_api_key) {\n    throw new Error('AI 客服未配置，请在后台设置中配置 AI 服务商和 API Key');\n  }\n\n  const baseUrl = config.ai_base_url || getProviderBaseUrl(config.ai_provider);\n  if (!baseUrl) throw new Error('未知的 AI 服务商: ' + config.ai_provider);\n\n  const model = config.ai_model || 'deepseek-chat';\n\n  // 构建 system prompt\n  let systemPrompt = config.ai_system_prompt || '你是 imai.work 的智能客服助手，专门解答关于养号、获客、短视频运营的问题。请用中文回答，语气友好专业。如果不确定答案，请诚实说明并建议用户联系人工客服。';\n\n  // RAG 检索相关知识\n  const knowledge = getKnowledgeContext(userMessage);\n  if (knowledge) {\n    systemPrompt += '\\n\\n' + knowledge;\n  }\n\n  // 构建消息列表\n  const messages = [{ role: 'system', content: systemPrompt }];\n\n  // 加入对话历史（最近 10 轮）\n  const recentHistory = history.slice(-20);\n  for (const msg of recentHistory) {\n    messages.push({ role: msg.role, content: msg.content });\n  }\n\n  // 加入用户新消息（支持图片）\n  if (imageUrl) {\n    messages.push({\n      role: 'user',\n      content: [\n        { type: 'text', text: userMessage || '请分析这张图片' },\n        { type: 'image_url', image_url: { url: imageUrl } },\n      ],\n    });\n  } else {\n    messages.push({ role: 'user', content: userMessage });\n  }\n\n  const reply = await callOpenAICompat(baseUrl, config.ai_api_key, model, messages);\n  return reply;\n}\n\nmodule.exports = { chat, getAIConfig };\n"}}}}