@astro-minimax/ai 0.9.0 → 0.9.3

package/dist/search/hybrid-search.d.ts

@@ -0,0 +1,111 @@
+/**
+ * RRF 混合检索模块
+ *
+ * 实现 Reciprocal Rank Fusion 算法，融合关键词检索和向量检索结果。
+ *
+ * RRF 公式: score(d) = Σ 1/(k + rank(d))
+ * 默认 k=60，源自 Cormack, Clarke, Buettcher 2009 论文。
+ */
+import type { ArticleContext } from "./types.js";
+export interface ArticleChunk {
+    id: string;
+    postId: string;
+    heading: string;
+    content: string;
+    position: number;
+    tokenCount: number;
+    headers: Record<string, string>;
+}
+export interface ArticleWithChunks extends ArticleContext {
+    chunks?: ArticleChunk[];
+}
+export interface RRFConfig {
+    /** RRF 常数 k，默认 60 */
+    k?: number;
+    /** BM25/TF-IDF 权重，默认 0.5 */
+    bm25Weight?: number;
+    /** 向量权重，默认 0.5 */
+    vectorWeight?: number;
+    /** 返回数量，默认 10 */
+    topK?: number;
+}
+export interface HybridSearchResult extends ArticleContext {
+    rrfScore: number;
+    bm25Rank?: number;
+    vectorRank?: number;
+    matchedChunks?: ArticleChunk[];
+}
+export interface ChunkMatchResult {
+    article: ArticleWithChunks;
+    chunk: ArticleChunk;
+    score: number;
+}
+export interface ChunkRelevanceOptions {
+    rawQuery?: string;
+    rawAnchors?: string[];
+}
+export interface NeighborChunkConfig {
+    includePrevious?: boolean;
+    includeNext?: boolean;
+    rawAnchors?: string[];
+}
+/**
+ * Reciprocal Rank Fusion 算法
+ *
+ * 融合多个检索系统的排名结果，只依赖排名而非分数。
+ *
+ * @param rankings - 多个检索系统的排名结果
+ * @param k - RRF 常数，默认 60
+ */
+export declare function reciprocalRankFusion(rankings: Array<Array<{
+    url: string;
+}>>, k?: number): Map<string, number>;
+/**
+ * 执行 BM25 + 向量混合检索
+ *
+ * @param query - 查询文本
+ * @param bm25Results - BM25/TF-IDF 检索结果
+ * @param vectorResults - 向量检索结果（可选）
+ * @param config - RRF 配置
+ */
+export declare function hybridSearch(_query: string, bm25Results: ArticleContext[], vectorResults: ArticleContext[] | null, config?: RRFConfig): HybridSearchResult[];
+/**
+ * 在段落级别搜索相关内容
+ *
+ * @param query - 查询文本
+ * @param articles - 包含段落的文章列表
+ * @param topK - 返回数量，默认 10
+ */
+export declare function searchChunks(query: string, articles: ArticleWithChunks[], topK?: number): ChunkMatchResult[];
+/**
+ * 计算段落相关性分数
+ *
+ * 考虑因素：
+ * 1. 标题匹配（权重 2.0）
+ * 2. 内容关键词匹配
+ * 3. 标题层级匹配（H1/H2 权重更高）
+ */
+export declare function computeChunkRelevance(queryTokens: string[], chunk: ArticleChunk, article?: Pick<ArticleContext, "title" | "categories" | "keyPoints">, options?: ChunkRelevanceOptions): number;
+/**
+ * 从段落匹配结果中格式化注入内容
+ *
+ * @param matches - 段落匹配结果
+ * @param maxTokens - 最大 token 数，默认 2000
+ */
+export declare function formatChunksForInjection(matches: ChunkMatchResult[], maxTokens?: number): string;
+export interface ChunkInjectionConfig {
+    /** 最大注入 token 数 */
+    maxTokens: number;
+    /** 最低相关性分数 */
+    minChunkScore: number;
+    /** 每篇文章最多注入段落数 */
+    maxChunksPerArticle: number;
+    rawAnchors?: string[];
+    currentArticleId?: string;
+}
+/**
+ * 选择最相关的段落用于注入
+ */
+export declare function selectRelevantChunks(query: string, articles: ArticleWithChunks[], config?: Partial<ChunkInjectionConfig>): ChunkMatchResult[];
+export declare function expandChunkMatchesWithNeighbors(matches: ChunkMatchResult[], config?: NeighborChunkConfig): ChunkMatchResult[];
+//# sourceMappingURL=hybrid-search.d.ts.map

package/dist/search/hybrid-search.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"hybrid-search.d.ts","sourceRoot":"","sources":["../../src/search/hybrid-search.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AAEH,OAAO,KAAK,EAAE,cAAc,EAAE,MAAM,YAAY,CAAC;AAKjD,MAAM,WAAW,YAAY;IAC3B,EAAE,EAAE,MAAM,CAAC;IACX,MAAM,EAAE,MAAM,CAAC;IACf,OAAO,EAAE,MAAM,CAAC;IAChB,OAAO,EAAE,MAAM,CAAC;IAChB,QAAQ,EAAE,MAAM,CAAC;IACjB,UAAU,EAAE,MAAM,CAAC;IACnB,OAAO,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;CACjC;AAED,MAAM,WAAW,iBAAkB,SAAQ,cAAc;IACvD,MAAM,CAAC,EAAE,YAAY,EAAE,CAAC;CACzB;AAED,MAAM,WAAW,SAAS;IACxB,qBAAqB;IACrB,CAAC,CAAC,EAAE,MAAM,CAAC;IACX,4BAA4B;IAC5B,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,kBAAkB;IAClB,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,iBAAiB;IACjB,IAAI,CAAC,EAAE,MAAM,CAAC;CACf;AAED,MAAM,WAAW,kBAAmB,SAAQ,cAAc;IACxD,QAAQ,EAAE,MAAM,CAAC;IACjB,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,aAAa,CAAC,EAAE,YAAY,EAAE,CAAC;CAChC;AAED,MAAM,WAAW,gBAAgB;IAC/B,OAAO,EAAE,iBAAiB,CAAC;IAC3B,KAAK,EAAE,YAAY,CAAC;IACpB,KAAK,EAAE,MAAM,CAAC;CACf;AAED,MAAM,WAAW,qBAAqB;IACpC,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB,UAAU,CAAC,EAAE,MAAM,EAAE,CAAC;CACvB;AAED,MAAM,WAAW,mBAAmB;IAClC,eAAe,CAAC,EAAE,OAAO,CAAC;IAC1B,WAAW,CAAC,EAAE,OAAO,CAAC;IACtB,UAAU,CAAC,EAAE,MAAM,EAAE,CAAC;CACvB;AASD;;;;;;;GAOG;AACH,wBAAgB,oBAAoB,CAClC,QAAQ,EAAE,KAAK,CAAC,KAAK,CAAC;IAAE,GAAG,EAAE,MAAM,CAAA;CAAE,CAAC,CAAC,EACvC,CAAC,GAAE,MAAsB,GACxB,GAAG,CAAC,MAAM,EAAE,MAAM,CAAC,CAYrB;AAID;;;;;;;GAOG;AACH,wBAAgB,YAAY,CAC1B,MAAM,EAAE,MAAM,EACd,WAAW,EAAE,cAAc,EAAE,EAC7B,aAAa,EAAE,cAAc,EAAE,GAAG,IAAI,EACtC,MAAM,CAAC,EAAE,SAAS,GACjB,kBAAkB,EAAE,CA8CtB;AAID;;;;;;GAMG;AACH,wBAAgB,YAAY,CAC1B,KAAK,EAAE,MAAM,EACb,QAAQ,EAAE,iBAAiB,EAAE,EAC7B,IAAI,GAAE,MAAW,GAChB,gBAAgB,EAAE,CAkBpB;AAED;;;;;;;GAOG;AACH,wBAAgB,qBAAqB,CACnC,WAAW,EAAE,MAAM,EAAE,EACrB,KAAK,EAAE,YAAY,EACnB,OAAO,CAAC,EAAE,IAAI,CAAC,cAAc,EAAE,OAAO,GAAG,YAAY,GAAG,WAAW,CAAC,EACpE,OAAO,GAAE,qBAA0B,GAClC,MAAM,CAoGR;AA2FD;;;;;GAKG;AACH,wBAAgB,wBAAwB,CACtC,OAAO,EAAE,gBAAgB,EAAE,EAC3B,SAAS,GAAE,MAAa,GACvB,MAAM,CAiBR;AAkBD,MAAM,WAAW,oBAAoB;IACnC,mBAAmB;IACnB,SAAS,EAAE,MAAM,CAAC;IAClB,cAAc;IACd,aAAa,EAAE,MAAM,CAAC;IACtB,kBAAkB;IAClB,mBAAmB,EAAE,MAAM,CAAC;IAC5B,UAAU,CAAC,EAAE,MAAM,EAAE,CAAC;IACtB,gBAAgB,CAAC,EAAE,MAAM,CAAC;CAC3B;AAQD;;GAEG;AACH,wBAAgB,oBAAoB,CAClC,KAAK,EAAE,MAAM,EACb,QAAQ,EAAE,iBAAiB,EAAE,EAC7B,MAAM,CAAC,EAAE,OAAO,CAAC,oBAAoB,CAAC,GACrC,gBAAgB,EAAE,CAmDpB;AAED,wBAAgB,+BAA+B,CAC7C,OAAO,EAAE,gBAAgB,EAAE,EAC3B,MAAM,GAAE,mBAAkE,GACzE,gBAAgB,EAAE,CAsDpB"}

package/dist/search/hybrid-search.js

@@ -0,0 +1,326 @@
+import { tokenize, normalizeText, extractCodeAnchors } from "../utils/text.js";
+const DEFAULT_RRF_K = 60;
+const DEFAULT_TOP_K = 10;
+function reciprocalRankFusion(rankings, k = DEFAULT_RRF_K) {
+  const scores = /* @__PURE__ */ new Map();
+  for (const ranking of rankings) {
+    for (let rank = 1; rank <= ranking.length; rank++) {
+      const doc = ranking[rank - 1];
+      const prev = scores.get(doc.url) ?? 0;
+      scores.set(doc.url, prev + 1 / (k + rank));
+    }
+  }
+  return new Map([...scores.entries()].sort((a, b) => b[1] - a[1]));
+}
+function hybridSearch(_query, bm25Results, vectorResults, config) {
+  const { k = DEFAULT_RRF_K, topK = DEFAULT_TOP_K } = config || {};
+  if (!bm25Results.length) return [];
+  if (!vectorResults || !vectorResults.length) {
+    return bm25Results.slice(0, topK).map((r, i) => ({
+      ...r,
+      rrfScore: 1 / (k + i + 1),
+      bm25Rank: i + 1
+    }));
+  }
+  const bm25Ranking = bm25Results.map((r) => ({ url: r.url }));
+  const vectorRanking = vectorResults.map((r) => ({ url: r.url }));
+  const rrfScores = reciprocalRankFusion([bm25Ranking, vectorRanking], k);
+  const bm25Map = new Map(bm25Results.map((r) => [r.url, r]));
+  const vectorMap = new Map(vectorResults.map((r, i) => [r.url, i + 1]));
+  const results = [];
+  let rank = 1;
+  for (const [url, rrfScore] of rrfScores) {
+    const bm25Result = bm25Map.get(url);
+    const vectorRank = vectorMap.get(url);
+    if (bm25Result) {
+      results.push({
+        ...bm25Result,
+        rrfScore,
+        bm25Rank: bm25Results.findIndex((r) => r.url === url) + 1,
+        vectorRank
+      });
+    }
+    if (results.length >= topK) break;
+    rank++;
+  }
+  return results;
+}
+function searchChunks(query, articles, topK = 10) {
+  const queryTokens = tokenize(query);
+  if (!queryTokens.length || !articles.length) return [];
+  const results = [];
+  for (const article of articles) {
+    if (!article.chunks?.length) continue;
+    for (const chunk of article.chunks) {
+      const score = computeChunkRelevance(queryTokens, chunk, article);
+      if (score > 0) {
+        results.push({ article, chunk, score });
+      }
+    }
+  }
+  return results.sort((a, b) => b.score - a.score).slice(0, topK);
+}
+function computeChunkRelevance(queryTokens, chunk, article, options = {}) {
+  let score = 0;
+  const rawQuery = options.rawQuery?.trim() ?? "";
+  const rawAnchors = options.rawAnchors && options.rawAnchors.length > 0 ? options.rawAnchors : extractCodeAnchors(rawQuery);
+  const headingTokens = tokenize(chunk.heading);
+  const contentTokens = tokenize(chunk.content);
+  const articleTitleTokens = article ? tokenize(article.title) : [];
+  const articleMetaText = article ? normalizeText(
+    [
+      article.title,
+      ...article.categories ?? [],
+      ...article.keyPoints ?? []
+    ].join(" ")
+  ) : "";
+  const anchorTokens = queryTokens.filter((token) => token.length >= 2).sort((a, b) => b.length - a.length).slice(0, 3);
+  const genericTopicTokens = /* @__PURE__ */ new Set([
+    "ai",
+    "rag",
+    "\u529F\u80FD",
+    "\u7279\u6027",
+    "\u529F\u80FD\u7279\u6027",
+    "\u8BBE\u8BA1",
+    "\u67B6\u6784",
+    "\u6A21\u5757",
+    "\u804A\u5929",
+    "\u914D\u7F6E"
+  ]);
+  const strongAnchorTokens = anchorTokens.filter(
+    (token) => !genericTopicTokens.has(token)
+  );
+  const headingText = normalizeText(chunk.heading);
+  const titleText = article ? normalizeText(article.title) : "";
+  score += scoreExactQueryMatches(rawQuery, chunk);
+  score += scoreExactCodeAnchorMatches(rawAnchors, chunk, article);
+  for (const token of queryTokens) {
+    if (headingTokens.some((h) => h.includes(token) || token.includes(h))) {
+      score += 2;
+    }
+    if (articleTitleTokens.some(
+      (title) => title.includes(token) || token.includes(title)
+    )) {
+      score += 1.2;
+    }
+  }
+  let contentMatches = 0;
+  for (const token of queryTokens) {
+    if (contentTokens.some((c) => c.includes(token) || token.includes(c))) {
+      contentMatches++;
+    }
+  }
+  score += contentMatches / queryTokens.length * 1.5;
+  if (chunk.headers.H1 || chunk.headers.H2) {
+    score *= 1.1;
+  }
+  if (anchorTokens.length > 0) {
+    const anchorHitCount = anchorTokens.filter(
+      (term) => articleMetaText.includes(term)
+    ).length;
+    if (anchorHitCount === 0) {
+      score *= 0.72;
+    } else {
+      score *= 1 + Math.min(anchorHitCount, 2) * 0.08;
+    }
+  }
+  if (strongAnchorTokens.length > 0) {
+    const strongTitleHits = strongAnchorTokens.filter(
+      (token) => titleText.includes(token)
+    ).length;
+    const strongHeadingHits = strongAnchorTokens.filter(
+      (token) => headingText.includes(token)
+    ).length;
+    if (strongTitleHits === 0 && strongHeadingHits === 0) {
+      score *= 0.58;
+    } else {
+      score *= 1 + Math.min(strongTitleHits + strongHeadingHits, 2) * 0.12;
+    }
+  }
+  return score;
+}
+function scoreExactCodeAnchorMatches(rawAnchors, chunk, article) {
+  if (rawAnchors.length === 0) return 0;
+  let bonus = 0;
+  const normalizedContent = normalizeText(chunk.content);
+  const normalizedHeading = normalizeText(chunk.heading);
+  const normalizedTitle = article ? normalizeText(article.title) : "";
+  for (const anchor of rawAnchors) {
+    if (anchor.length < 2) continue;
+    const normalizedAnchor = normalizeText(anchor);
+    if (!normalizedAnchor) continue;
+    if (chunk.content.includes(anchor)) {
+      bonus += 5.5;
+      continue;
+    }
+    if (chunk.heading.includes(anchor)) {
+      bonus += 4.2;
+      continue;
+    }
+    if (article?.title.includes(anchor)) {
+      bonus += 2.5;
+      continue;
+    }
+    if (normalizedContent.includes(normalizedAnchor)) {
+      bonus += 3.5;
+    } else if (normalizedHeading.includes(normalizedAnchor)) {
+      bonus += 2.2;
+    } else if (normalizedTitle.includes(normalizedAnchor)) {
+      bonus += 1.4;
+    }
+  }
+  return bonus;
+}
+function scoreExactQueryMatches(query, chunk) {
+  const candidate = extractLikelyQuotedText(query);
+  if (!candidate) return 0;
+  const normalizedCandidate = normalizeText(candidate);
+  if (normalizedCandidate.length < 12) return 0;
+  const chunkContent = chunk.content;
+  const normalizedContent = normalizeText(chunkContent);
+  const normalizedHeading = normalizeText(chunk.heading);
+  let bonus = 0;
+  if (chunkContent.includes(candidate)) {
+    bonus += 6;
+  } else if (normalizedContent.includes(normalizedCandidate)) {
+    bonus += 4.5;
+  }
+  if (!chunk.heading && normalizedContent.startsWith(normalizedCandidate)) {
+    bonus += 8;
+  }
+  if (normalizedHeading.includes(normalizedCandidate)) {
+    bonus += 1.5;
+  }
+  return bonus;
+}
+function extractLikelyQuotedText(query) {
+  const trimmed = query.trim();
+  if (!trimmed) return "";
+  const quotedMatches = [...trimmed.matchAll(/["“”'‘’「」『』《》](.+?)["“”'‘’「」『』《》]/g)].map((match) => match[1]?.trim() ?? "").filter(Boolean).sort((a, b) => b.length - a.length);
+  if (quotedMatches[0]) {
+    return quotedMatches[0];
+  }
+  return trimmed;
+}
+function formatChunksForInjection(matches, maxTokens = 2e3) {
+  if (!matches.length) return "";
+  const lines = [];
+  let totalTokens = 0;
+  for (const match of matches) {
+    const chunkText = formatChunkForInjection(match);
+    const chunkTokens = estimateTokens(chunkText);
+    if (totalTokens + chunkTokens > maxTokens) break;
+    lines.push(chunkText);
+    totalTokens += chunkTokens;
+  }
+  return lines.join("\n\n");
+}
+function formatChunkForInjection(match) {
+  const { article, chunk } = match;
+  const heading = chunk.heading ? `\u3010${chunk.heading}\u3011` : "";
+  const source = `\u6765\u6E90: [${article.title}](${article.url})`;
+  return `${heading}
+${chunk.content.slice(0, 500)}
+
+${source}`;
+}
+function estimateTokens(text) {
+  const cjkChars = (text.match(/[\u4e00-\u9fff\u3400-\u4dbf]/g) || []).length;
+  const nonCjkChars = text.length - cjkChars;
+  return Math.ceil(cjkChars / 2) + Math.ceil(nonCjkChars / 4);
+}
+const DEFAULT_INJECTION_CONFIG = {
+  maxTokens: 2e3,
+  minChunkScore: 0.3,
+  maxChunksPerArticle: 3
+};
+function selectRelevantChunks(query, articles, config) {
+  const cfg = { ...DEFAULT_INJECTION_CONFIG, ...config };
+  const queryTokens = tokenize(query);
+  if (!queryTokens.length) return [];
+  const allMatches = [];
+  for (const article of articles) {
+    if (!article.chunks?.length) continue;
+    const articleMatches = [];
+    for (const chunk of article.chunks) {
+      const score = computeChunkRelevance(queryTokens, chunk, article, {
+        rawQuery: query,
+        rawAnchors: cfg.rawAnchors
+      });
+      if (score >= cfg.minChunkScore) {
+        articleMatches.push({ article, chunk, score });
+      }
+    }
+    articleMatches.sort((a, b) => b.score - a.score);
+    allMatches.push(...articleMatches.slice(0, cfg.maxChunksPerArticle));
+  }
+  const globallyRanked = allMatches.sort((a, b) => b.score - a.score);
+  const selected = cfg.currentArticleId ? [
+    ...globallyRanked.filter((match) => match.article.id === cfg.currentArticleId),
+    ...globallyRanked.filter((match) => match.article.id !== cfg.currentArticleId)
+  ].slice(0, 20) : globallyRanked.slice(0, 20);
+  log.debug(
+    `selectRelevantChunks: queryTokens=${queryTokens.length}, articles=${articles.length}, matched=${allMatches.length}, selected=${selected.length}, maxPerArticle=${cfg.maxChunksPerArticle}, minScore=${cfg.minChunkScore}`
+  );
+  if (selected.length > 0) {
+    log.debug(
+      `selectRelevantChunks top: ${selected.slice(0, 5).map(
+        (match) => `${match.article.title}#${match.chunk.id}:${match.score.toFixed(3)}`
+      ).join(", ")}`
+    );
+  }
+  return selected;
+}
+function expandChunkMatchesWithNeighbors(matches, config = { includePrevious: true, includeNext: true }) {
+  const expanded = [];
+  const seen = /* @__PURE__ */ new Set();
+  const rawAnchors = config.rawAnchors ?? [];
+  const candidateHasAnchor = (candidate) => {
+    if (rawAnchors.length === 0) return true;
+    return rawAnchors.some((anchor) => {
+      const normalizedAnchor = normalizeText(anchor);
+      return candidate.chunk.content.includes(anchor) || candidate.chunk.heading.includes(anchor) || normalizeText(candidate.chunk.content).includes(normalizedAnchor) || normalizeText(candidate.chunk.heading).includes(normalizedAnchor);
+    });
+  };
+  for (const match of matches) {
+    const push = (candidate) => {
+      if (!candidate) return;
+      if (seen.has(candidate.chunk.id)) return;
+      if (candidate !== match && !candidateHasAnchor(candidate)) return;
+      seen.add(candidate.chunk.id);
+      expanded.push(candidate);
+    };
+    push(match);
+    const articleChunks = match.article.chunks ?? [];
+    const index = articleChunks.findIndex((chunk) => chunk.id === match.chunk.id);
+    if (index === -1) continue;
+    if (config.includePrevious && index > 0) {
+      push({
+        article: match.article,
+        chunk: articleChunks[index - 1],
+        score: match.score * 0.85
+      });
+    }
+    if (config.includeNext && index < articleChunks.length - 1) {
+      push({
+        article: match.article,
+        chunk: articleChunks[index + 1],
+        score: match.score * 0.9
+      });
+    }
+  }
+  return expanded.sort(
+    (a, b) => b.score - a.score || a.chunk.position - b.chunk.position
+  );
+}
+import { createLogger } from "../utils/logger.js";
+const log = createLogger("hybrid-search");
+export {
+  computeChunkRelevance,
+  expandChunkMatchesWithNeighbors,
+  formatChunksForInjection,
+  hybridSearch,
+  reciprocalRankFusion,
+  searchChunks,
+  selectRelevantChunks
+};

package/dist/search/index.d.ts

@@ -1,10 +1,12 @@
-export { initArticleIndex, initProjectIndex, searchArticles, searchProjects, mergeResults } from './search-api.js';
-export { getIDFMapForIndex } from './search-index.js';
-export { loadVectorIndex, clearVectorIndex, hasVectorIndex, rerankWithVectors } from './vector-reranker.js';
-export type { VectorIndex, VectorChunk } from './vector-reranker.js';
-export { getSessionCacheKey, getCachedContext, setCachedContext, deleteCachedContext, setCacheAdapter, getCacheAdapter, cleanupCache, SESSION_CACHE_TTL_SECONDS, SESSION_CACHE_TTL_MS, getCachedContextSync, setCachedContextSync, cleanupCacheLegacy, } from './session-cache.js';
-export { normalizeText, tokenize, scoreDocument } from './search-utils.js';
-export { buildIDFMap, getIDFWeight } from './idf.js';
-export type { IDFMap } from './idf.js';
-export type { SearchDocument, ArticleContext, ProjectContext, CachedSearchContext, SearchResult } from './types.js';
+export { initArticleIndex, initProjectIndex, initArticleChunks, hasArticleChunks, getArticleChunks, searchArticles, searchProjects, searchArticleChunks, mergeResults, } from "./search-api.js";
+export { getIDFMapForIndex } from "./search-index.js";
+export { loadVectorIndex, clearVectorIndex, hasVectorIndex, rerankWithVectors, } from "./vector-reranker.js";
+export type { VectorIndex, VectorChunk } from "./vector-reranker.js";
+export { getSessionCacheKey, getCachedContext, setCachedContext, deleteCachedContext, setCacheAdapter, getCacheAdapter, cleanupCache, SESSION_CACHE_TTL_SECONDS, SESSION_CACHE_TTL_MS, } from "./session-cache.js";
+export { normalizeText, tokenize } from "../utils/text.js";
+export { scoreDocument, filterLowRelevance, pickAnchorTerms, } from "./scoring.js";
+export { buildIDFMap, getIDFWeight } from "./idf.js";
+export type { IDFMap } from "./idf.js";
+export type { SearchDocument, ArticleContext, ArticleChunk, ProjectContext, CachedSearchContext, SearchResult, } from "./types.js";
+export { reciprocalRankFusion, hybridSearch, searchChunks, computeChunkRelevance, selectRelevantChunks, expandChunkMatchesWithNeighbors, formatChunksForInjection, type RRFConfig, type HybridSearchResult, type ChunkMatchResult, type ChunkInjectionConfig, } from "./hybrid-search.js";
 //# sourceMappingURL=index.d.ts.map

package/dist/search/index.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../src/search/index.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,gBAAgB,EAAE,gBAAgB,EAAE,cAAc,EAAE,cAAc,EAAE,YAAY,EAAE,MAAM,iBAAiB,CAAC;AACnH,OAAO,EAAE,iBAAiB,EAAE,MAAM,mBAAmB,CAAC;AACtD,OAAO,EAAE,eAAe,EAAE,gBAAgB,EAAE,cAAc,EAAE,iBAAiB,EAAE,MAAM,sBAAsB,CAAC;AAC5G,YAAY,EAAE,WAAW,EAAE,WAAW,EAAE,MAAM,sBAAsB,CAAC;AACrE,OAAO,EACL,kBAAkB,EAClB,gBAAgB,EAChB,gBAAgB,EAChB,mBAAmB,EACnB,eAAe,EACf,eAAe,EACf,YAAY,EACZ,yBAAyB,EACzB,oBAAoB,EACpB,oBAAoB,EACpB,oBAAoB,EACpB,kBAAkB,GACnB,MAAM,oBAAoB,CAAC;AAC5B,OAAO,EAAE,aAAa,EAAE,QAAQ,EAAE,aAAa,EAAE,MAAM,mBAAmB,CAAC;AAC3E,OAAO,EAAE,WAAW,EAAE,YAAY,EAAE,MAAM,UAAU,CAAC;AACrD,YAAY,EAAE,MAAM,EAAE,MAAM,UAAU,CAAC;AACvC,YAAY,EAAE,cAAc,EAAE,cAAc,EAAE,cAAc,EAAE,mBAAmB,EAAE,YAAY,EAAE,MAAM,YAAY,CAAC"}
+{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../src/search/index.ts"],"names":[],"mappings":"AAAA,OAAO,EACL,gBAAgB,EAChB,gBAAgB,EAChB,iBAAiB,EACjB,gBAAgB,EAChB,gBAAgB,EAChB,cAAc,EACd,cAAc,EACd,mBAAmB,EACnB,YAAY,GACb,MAAM,iBAAiB,CAAC;AACzB,OAAO,EAAE,iBAAiB,EAAE,MAAM,mBAAmB,CAAC;AACtD,OAAO,EACL,eAAe,EACf,gBAAgB,EAChB,cAAc,EACd,iBAAiB,GAClB,MAAM,sBAAsB,CAAC;AAC9B,YAAY,EAAE,WAAW,EAAE,WAAW,EAAE,MAAM,sBAAsB,CAAC;AACrE,OAAO,EACL,kBAAkB,EAClB,gBAAgB,EAChB,gBAAgB,EAChB,mBAAmB,EACnB,eAAe,EACf,eAAe,EACf,YAAY,EACZ,yBAAyB,EACzB,oBAAoB,GACrB,MAAM,oBAAoB,CAAC;AAC5B,OAAO,EAAE,aAAa,EAAE,QAAQ,EAAE,MAAM,kBAAkB,CAAC;AAC3D,OAAO,EACL,aAAa,EACb,kBAAkB,EAClB,eAAe,GAChB,MAAM,cAAc,CAAC;AACtB,OAAO,EAAE,WAAW,EAAE,YAAY,EAAE,MAAM,UAAU,CAAC;AACrD,YAAY,EAAE,MAAM,EAAE,MAAM,UAAU,CAAC;AACvC,YAAY,EACV,cAAc,EACd,cAAc,EACd,YAAY,EACZ,cAAc,EACd,mBAAmB,EACnB,YAAY,GACb,MAAM,YAAY,CAAC;AACpB,OAAO,EACL,oBAAoB,EACpB,YAAY,EACZ,YAAY,EACZ,qBAAqB,EACrB,oBAAoB,EACpB,+BAA+B,EAC/B,wBAAwB,EACxB,KAAK,SAAS,EACd,KAAK,kBAAkB,EACvB,KAAK,gBAAgB,EACrB,KAAK,oBAAoB,GAC1B,MAAM,oBAAoB,CAAC"}

package/dist/search/index.js

@@ -1,6 +1,21 @@
-import { initArticleIndex, initProjectIndex, searchArticles, searchProjects, mergeResults } from "./search-api.js";
+import {
+  initArticleIndex,
+  initProjectIndex,
+  initArticleChunks,
+  hasArticleChunks,
+  getArticleChunks,
+  searchArticles,
+  searchProjects,
+  searchArticleChunks,
+  mergeResults
+} from "./search-api.js";
 import { getIDFMapForIndex } from "./search-index.js";
-import { loadVectorIndex, clearVectorIndex, hasVectorIndex, rerankWithVectors } from "./vector-reranker.js";
+import {
+  loadVectorIndex,
+  clearVectorIndex,
+  hasVectorIndex,
+  rerankWithVectors
+} from "./vector-reranker.js";
 import {
   getSessionCacheKey,
   getCachedContext,
@@ -10,39 +25,60 @@ import {
   getCacheAdapter,
   cleanupCache,
   SESSION_CACHE_TTL_SECONDS,
-  SESSION_CACHE_TTL_MS,
-  getCachedContextSync,
-  setCachedContextSync,
-  cleanupCacheLegacy
+  SESSION_CACHE_TTL_MS
 } from "./session-cache.js";
-import { normalizeText, tokenize, scoreDocument } from "./search-utils.js";
+import { normalizeText, tokenize } from "../utils/text.js";
+import {
+  scoreDocument,
+  filterLowRelevance,
+  pickAnchorTerms
+} from "./scoring.js";
 import { buildIDFMap, getIDFWeight } from "./idf.js";
+import {
+  reciprocalRankFusion,
+  hybridSearch,
+  searchChunks,
+  computeChunkRelevance,
+  selectRelevantChunks,
+  expandChunkMatchesWithNeighbors,
+  formatChunksForInjection
+} from "./hybrid-search.js";
 export {
   SESSION_CACHE_TTL_MS,
   SESSION_CACHE_TTL_SECONDS,
   buildIDFMap,
   cleanupCache,
-  cleanupCacheLegacy,
   clearVectorIndex,
+  computeChunkRelevance,
   deleteCachedContext,
+  expandChunkMatchesWithNeighbors,
+  filterLowRelevance,
+  formatChunksForInjection,
+  getArticleChunks,
   getCacheAdapter,
   getCachedContext,
-  getCachedContextSync,
   getIDFMapForIndex,
   getIDFWeight,
   getSessionCacheKey,
+  hasArticleChunks,
   hasVectorIndex,
+  hybridSearch,
+  initArticleChunks,
   initArticleIndex,
   initProjectIndex,
   loadVectorIndex,
   mergeResults,
   normalizeText,
+  pickAnchorTerms,
+  reciprocalRankFusion,
   rerankWithVectors,
   scoreDocument,
+  searchArticleChunks,
   searchArticles,
+  searchChunks,
   searchProjects,
+  selectRelevantChunks,
   setCacheAdapter,
   setCachedContext,
-  setCachedContextSync,
   tokenize
 };

package/dist/search/scoring.d.ts

@@ -0,0 +1,18 @@
+import type { IDFMap } from "./idf.js";
+export declare function scoreDocument(tokens: string[], doc: {
+    title: string;
+    content: string;
+    excerpt: string;
+    keyPoints: string[];
+    categories: string[];
+    tags: string[];
+}, idfMap?: IDFMap | null): number;
+export declare function filterLowRelevance<T extends {
+    score: number;
+}>(results: T[], relativeThreshold?: number, minAbsoluteScore?: number): T[];
+export declare function pickAnchorTerms(query: string, candidates: Array<{
+    title: string;
+    keyPoints: string[];
+    categories: string[];
+}>, maxTerms?: number, minTermLength?: number): string[];
+//# sourceMappingURL=scoring.d.ts.map

package/dist/search/scoring.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"scoring.d.ts","sourceRoot":"","sources":["../../src/search/scoring.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,MAAM,EAAE,MAAM,UAAU,CAAC;AAavC,wBAAgB,aAAa,CAC3B,MAAM,EAAE,MAAM,EAAE,EAChB,GAAG,EAAE;IACH,KAAK,EAAE,MAAM,CAAC;IACd,OAAO,EAAE,MAAM,CAAC;IAChB,OAAO,EAAE,MAAM,CAAC;IAChB,SAAS,EAAE,MAAM,EAAE,CAAC;IACpB,UAAU,EAAE,MAAM,EAAE,CAAC;IACrB,IAAI,EAAE,MAAM,EAAE,CAAC;CAChB,EACD,MAAM,CAAC,EAAE,MAAM,GAAG,IAAI,GACrB,MAAM,CAwBR;AAED,wBAAgB,kBAAkB,CAAC,CAAC,SAAS;IAAE,KAAK,EAAE,MAAM,CAAA;CAAE,EAC5D,OAAO,EAAE,CAAC,EAAE,EACZ,iBAAiB,SAAO,EACxB,gBAAgB,SAAI,GACnB,CAAC,EAAE,CAUL;AAED,wBAAgB,eAAe,CAC7B,KAAK,EAAE,MAAM,EACb,UAAU,EAAE,KAAK,CAAC;IAChB,KAAK,EAAE,MAAM,CAAC;IACd,SAAS,EAAE,MAAM,EAAE,CAAC;IACpB,UAAU,EAAE,MAAM,EAAE,CAAC;CACtB,CAAC,EACF,QAAQ,SAAI,EACZ,aAAa,SAAI,GAChB,MAAM,EAAE,CA2BV"}

package/dist/search/{search-utils.js → scoring.js}

@@ -1,22 +1,5 @@
 import { getIDFWeight } from "./idf.js";
-function normalizeText(text) {
-  return text.toLowerCase().replace(/[^\u4e00-\u9fa5\w\s]/g, " ").replace(/\s+/g, " ").trim();
-}
-function tokenize(text) {
-  const normalized = normalizeText(text);
-  const parts = normalized.split(/\s+/).filter(Boolean);
-  return dedupeByContainment(parts);
-}
-function dedupeByContainment(terms) {
-  const unique = [...new Set(terms)];
-  const kept = [];
-  for (const term of unique.sort((a, b) => b.length - a.length)) {
-    if (!kept.some((existing) => existing.includes(term))) {
-      kept.push(term);
-    }
-  }
-  return kept;
-}
+import { normalizeText, tokenize } from "../utils/text.js";
 const FIELD_WEIGHTS = {
   title: 8,
   keyPoints: 5,
@@ -47,11 +30,14 @@ function scoreDocument(tokens, doc, idfMap) {
   return score;
 }
 function filterLowRelevance(results, relativeThreshold = 0.35, minAbsoluteScore = 2) {
-  if (results.length <= 3) return results;
+  if (results.length <= 2) return results;
   const topScore = results[0]?.score ?? 0;
   if (topScore <= 0) return results;
-  const threshold = Math.max(minAbsoluteScore, topScore * relativeThreshold);
-  return results.filter((item, index) => index < 3 || item.score >= threshold);
+  const threshold = Math.max(
+    minAbsoluteScore,
+    topScore * (results.length > 8 ? relativeThreshold + 0.1 : relativeThreshold)
+  );
+  return results.filter((item, index) => index < 2 || item.score >= threshold);
 }
 function pickAnchorTerms(query, candidates, maxTerms = 2, minTermLength = 2) {
   const terms = tokenize(query).filter((t) => t.length >= minTermLength);
@@ -59,8 +45,12 @@ function pickAnchorTerms(query, candidates, maxTerms = 2, minTermLength = 2) {
   if (!candidates.length) return terms.slice(0, maxTerms);
   const scored = terms.map((term) => {
     let hitCount = 0;
-    for (const c of candidates) {
-      const text = normalizeText([c.title, ...c.keyPoints, ...c.categories].join(" "));
+    for (const candidate of candidates) {
+      const text = normalizeText(
+        [candidate.title, ...candidate.keyPoints, ...candidate.categories].join(
+          " "
+        )
+      );
       if (text.includes(term)) hitCount++;
     }
     if (hitCount <= 0) return { term, score: Number.NEGATIVE_INFINITY };
@@ -72,10 +62,7 @@ function pickAnchorTerms(query, candidates, maxTerms = 2, minTermLength = 2) {
   return scored.filter((s) => Number.isFinite(s.score)).sort((a, b) => b.score - a.score).map((s) => s.term).slice(0, maxTerms);
 }
 export {
-  dedupeByContainment,
   filterLowRelevance,
-  normalizeText,
   pickAnchorTerms,
-  scoreDocument,
-  tokenize
+  scoreDocument
 };