@chiway/contextweaver 1.4.0 → 1.5.0

package/README.md

@@ -24,10 +24,11 @@
 - **RRF 融合 (Reciprocal Rank Fusion)**：智能融合多路召回结果
 
 ### 🧠 AST 语义分片
-- **Tree-sitter 解析**：支持 TypeScript、JavaScript、Python、Go、Java、Rust 六大语言
+- **Tree-sitter 解析**：支持 TypeScript、JavaScript、Python、Go、Java、Rust、C、C++、C# 等语言
 - **Dual-Text 策略**：`displayCode` 用于展示，`vectorText` 用于 Embedding
 - **Gap-Aware 合并**：智能处理代码间隙，保持语义完整性
 - **Breadcrumb 注入**：向量文本包含层级路径，提升检索召回率
+- **UTF-16 字符域归一**：在写入 metadata 前用 `SourceAdapter.toCharOffset` 统一偏移，避免多字节字符切片错位（v1.4.0+）
 
 ### 📊 三阶段上下文扩展
 - **E1 邻居扩展**：同文件前后相邻 chunks，保证代码块完整性
@@ -44,6 +45,13 @@
 - **意图与术语分离**：LLM 友好的 API 设计
 - **自动索引**：首次查询自动触发索引，增量更新透明无感
 
+### 🛡️ Crash-Safe 数据架构 (v1.4.0+)
+- **正文唯一源**：LanceDB 仅存向量与定位元数据，正文回查 `files.content`，索引体积降低 30-50%
+- **跨库事务补偿**：LanceDB → FTS+outbox → SQLite mark 三阶段写入，任一失败自动回滚或重放
+- **迁移状态机**：`pending/done/aborted` 三态持久化，崩溃恢复自动重建
+- **跨进程互斥**：advisory lock 防止 MCP server 与 CLI 并发触发 LanceDB 迁移
+- **chunk_id 去重**：写入前预删除，防止 retry 场景产生重复行
+
 ## 📦 快速开始
 
 ### 环境要求
@@ -55,10 +63,10 @@
 
 ```bash
 # 全局安装
-npm install -g @hsingjui/contextweaver
+npm install -g @chiway/contextweaver
 
 # 或使用 pnpm
-pnpm add -g @hsingjui/contextweaver
+pnpm add -g @chiway/contextweaver
 ```
 
 ### 初始化配置
@@ -120,6 +128,20 @@ cw search --information-request "数据库连接逻辑" --technical-terms "Datab
 contextweaver mcp
 ```
 
+### 索引管理 (v1.4.0+)
+
+```bash
+# 查看 LanceDB 迁移状态
+contextweaver migrate
+
+# 解除 aborted 状态：清空 LanceDB 并触发全量重建
+# 触发时机：抽样校验失败后 Indexer 拒绝写入；运行此命令后再次 index 即可
+contextweaver migrate --reset
+
+# 指定项目路径
+contextweaver migrate --path /path/to/project
+```
+
 ## 🔧 MCP 集成配置
 
 ### Claude Desktop 配置
@@ -201,55 +223,93 @@ flowchart TB
 | **SearchService** | 混合搜索核心，协调向量/词法召回、RRF 融合、Rerank 精排 |
 | **GraphExpander** | 上下文扩展器，执行 E1/E2/E3 三阶段扩展策略 |
 | **ContextPacker** | 上下文打包器，负责段落合并和 Token 预算控制 |
-| **VectorStore** | LanceDB 适配层，管理向量索引的增删改查 |
-| **SQLite (FTS5)** | 元数据存储 + 全文搜索索引 |
-| **SemanticSplitter** | AST 语义分片器，基于 Tree-sitter 解析 |
+| **ChunkContentLoader** | 按 `(path, start_index, end_index)` 从 `files.content` 批量切片（v1.4.0+） |
+| **VectorStore** | LanceDB 适配层，仅暴露纯 vector 操作 |
+| **Database (SQLite)** | 元数据存储 + FTS5 全文索引，schema_version=3 |
+| **Bootstrap** | 跨库初始化协调器：pending_marks 重放 + LanceDB schema 迁移（v1.4.0+） |
+| **SemanticSplitter** | AST 语义分片器，基于 Tree-sitter 解析，写入时统一到 UTF-16 字符域 |
+
+### 数据架构 (v1.4.0+)
+
+```
+~/.contextweaver/<projectId>/
+├── index.db                 # SQLite
+│   ├── files                # 文件元数据 + 完整正文（content 列，文本切片唯一来源）
+│   ├── files_fts            # 外部内容表，倒排索引指向 files
+│   ├── chunks_fts           # chunk 级倒排索引，per-file 整体替换
+│   ├── metadata             # schema_version / lancedb_migration_state / lock
+│   └── pending_marks        # outbox：vector_index_hash 标记失败时启动重放
+└── vectors.lance/           # LanceDB chunks 表（仅向量 + 定位元数据，不存正文）
+```
+
+**关键不变量**：
+- 正文唯一源是 `files.content`；`ChunkContentLoader` 用 `start_index/end_index` 切片（与 `displayCode` 同源）
+- 所有 LanceDB 偏移字段都在 UTF-16 字符域，多字节文件不会切错
+- 跨库写入顺序：LanceDB → (FTS + outbox 单事务) → SQLite mark + 清 outbox
+- LanceDB 迁移状态 `pending/done/aborted` 持久化，跨进程用 advisory lock 互斥
 
 ## 📁 项目结构
 
 ```
 contextweaver/
 ├── src/
-│   ├── index.ts              # CLI 入口
+│   ├── index.ts              # CLI 入口（init / index / search / mcp / migrate）
 │   ├── config.ts             # 配置管理（环境变量）
 │   ├── api/                  # 外部 API 封装
-│   │   ├── embed.ts          # Embedding API
-│   │   └── rerank.ts         # Reranker API
+│   │   ├── embedding.ts      # Embedding API
+│   │   └── reranker.ts       # Reranker API
 │   ├── chunking/             # 语义分片
 │   │   ├── SemanticSplitter.ts   # AST 语义分片器
-│   │   ├── SourceAdapter.ts      # 源码适配器
+│   │   ├── SourceAdapter.ts      # 源码适配器（UTF-16/UTF-8 域归一）
 │   │   ├── LanguageSpec.ts       # 语言规范定义
-│   │   └── ParserPool.ts         # Tree-sitter 解析器池
+│   │   ├── ParserPool.ts         # Tree-sitter 解析器池
+│   │   └── types.ts              # 分片类型定义
 │   ├── scanner/              # 文件扫描
 │   │   ├── crawler.ts        # 文件系统遍历
 │   │   ├── processor.ts      # 文件处理
-│   │   └── filter.ts         # 过滤规则
+│   │   ├── filter.ts         # 过滤规则
+│   │   ├── hash.ts           # 文件 hash
+│   │   └── language.ts       # 语言识别
 │   ├── indexer/              # 索引器
-│   │   └── index.ts          # 批量索引逻辑
+│   │   └── index.ts          # 三阶段事务（LanceDB → FTS+outbox → SQLite mark）
 │   ├── vectorStore/          # 向量存储
-│   │   └── index.ts          # LanceDB 适配层
+│   │   └── index.ts          # LanceDB 适配层（纯 vector 操作）
 │   ├── db/                   # 数据库
-│   │   └── index.ts          # SQLite + FTS5
+│   │   ├── index.ts          # SQLite + FTS5 + pending_marks + 迁移状态机
+│   │   └── bootstrap.ts      # 跨库初始化协调（v1.4.0+）
 │   ├── search/               # 搜索服务
-│   │   ├── SearchService.ts  # 核心搜索服务
-│   │   ├── GraphExpander.ts  # 上下文扩展器
-│   │   ├── ContextPacker.ts  # 上下文打包器
-│   │   ├── fts.ts            # 全文搜索
-│   │   ├── config.ts         # 搜索配置
-│   │   ├── types.ts          # 类型定义
-│   │   └── resolvers/        # 多语言 Import 解析器
+│   │   ├── SearchService.ts      # 核心搜索服务
+│   │   ├── GraphExpander.ts      # 上下文扩展器
+│   │   ├── ContextPacker.ts      # 上下文打包器
+│   │   ├── ChunkContentLoader.ts # 按 (path, start_index, end_index) 切片（v1.4.0+）
+│   │   ├── fts.ts                # 全文搜索（per-file 整体替换）
+│   │   ├── config.ts             # 搜索配置
+│   │   ├── types.ts              # 类型定义
+│   │   ├── utils.ts              # token overlap 评分
+│   │   └── resolvers/            # 多语言 Import 解析器
 │   │       ├── JsTsResolver.ts
 │   │       ├── PythonResolver.ts
 │   │       ├── GoResolver.ts
 │   │       ├── JavaResolver.ts
-│   │       └── RustResolver.ts
+│   │       ├── RustResolver.ts
+│   │       ├── CppResolver.ts
+│   │       └── CSharpResolver.ts
 │   ├── mcp/                  # MCP 服务端
 │   │   ├── server.ts         # MCP 服务器实现
 │   │   ├── main.ts           # MCP 入口
 │   │   └── tools/
 │   │       └── codebaseRetrieval.ts  # 代码检索工具
 │   └── utils/                # 工具函数
-│       └── logger.ts         # 日志系统
+│       ├── logger.ts         # 日志系统
+│       ├── encoding.ts       # 编码检测
+│       └── lock.ts           # 文件锁
+├── tests/                    # 单测 + 集成测试（109 测试用例）
+│   ├── chunking/             # SourceAdapter / 分片
+│   ├── db/                   # 迁移、outbox、advisory lock
+│   ├── indexer/              # 事务补偿、GC、aborted 守卫
+│   ├── integration/          # 真实 LanceDB 端到端
+│   ├── search/               # FTS、ChunkContentLoader、Packer
+│   └── vectorStore/          # chunk_id 去重、抽样校验
 ├── package.json
 └── tsconfig.json
 ```
@@ -314,11 +374,14 @@ ContextWeaver 通过 Tree-sitter 原生支持以下编程语言的 AST 解析：
 | 语言 | AST 解析 | Import 解析 | 文件扩展名 |
 |------|----------|-------------|-----------|
 | TypeScript | ✅ | ✅ | `.ts`, `.tsx` |
-| JavaScript | ✅ | ✅ | `.js`, `.jsx`, `.mjs` |
+| JavaScript | ✅ | ✅ | `.js`, `.jsx`, `.mjs`, `.cjs` |
 | Python | ✅ | ✅ | `.py` |
 | Go | ✅ | ✅ | `.go` |
 | Java | ✅ | ✅ | `.java` |
 | Rust | ✅ | ✅ | `.rs` |
+| C | ✅ | ✅ | `.c`, `.h` |
+| C++ | ✅ | ✅ | `.cpp`, `.cc`, `.cxx`, `.hpp` |
+| C# | ✅ | ✅ | `.cs` |
 
 其他语言会采用基于行的 Fallback 分片策略，仍可正常索引和搜索。
 
@@ -327,11 +390,16 @@ ContextWeaver 通过 Tree-sitter 原生支持以下编程语言的 AST 解析：
 ### 索引流程
 
 ```
+0. Bootstrap   → pending_marks 重放 + LanceDB schema 迁移（首次启动）
 1. Crawler     → 遍历文件系统，过滤忽略项
 2. Processor   → 读取文件内容，计算 hash
-3. Splitter    → AST 解析，语义分片
-4. Indexer     → 批量 Embedding，写入向量库
-5. FTS Index   → 更新全文搜索索引
+3. Splitter    → AST 解析，语义分片（偏移归一到 UTF-16 字符域）
+4. Indexer     → 批量 Embedding
+5. 阶段 4-6 伪事务：
+   ├─ LanceDB 写入（预删 (path, hash) 防重复 → add → 清旧版本）
+   ├─ FTS + outbox 单 SQLite 事务（失败回滚 LanceDB）
+   └─ SQLite mark + 清 outbox 单事务（失败时 outbox 保留，下次启动 replay）
+6. 末尾 GC     → 清理 LanceDB 孤儿 chunks（time budget 5s）
 ```
 
 ### 搜索流程
@@ -366,6 +434,48 @@ ContextWeaver 通过 Tree-sitter 原生支持以下编程语言的 AST 解析：
 LOG_LEVEL=debug contextweaver search --information-request "..."
 ```
 
+## 🚨 故障排查 (v1.4.0+)
+
+### LanceDB 迁移卡死 (`aborted` 状态)
+
+**现象**：`contextweaver index` 报错 "LanceDB 处于 aborted 状态，拒绝写入以防止 schema 污染"。
+
+**原因**：v1.4.0 升级时 LanceDB 旧索引中的 `display_code` 与当前 `files.content` 抽样差异 > 1%（通常发生在 chunk 偏移用 UTF-8 字节域旧索引上）。
+
+**解决**：
+```bash
+contextweaver migrate --reset   # 清空 LanceDB chunks 表 + 重置状态为 done
+contextweaver index             # 全量重建（新 schema）
+```
+
+### 跨进程迁移竞争
+
+如果 MCP server 长驻 + 另一终端跑 `contextweaver index`，两进程会争抢迁移。v1.4.0 引入 10 分钟僵尸阈值的 advisory lock，自动让一个进程跳过迁移、另一个完成。
+
+如锁卡住（process kill -9 后），可手动清理：
+```bash
+sqlite3 ~/.contextweaver/<projectId>/index.db \
+  "DELETE FROM metadata WHERE key = 'lancedb_migration_lock';"
+```
+
+### 重复 embedding 浪费
+
+v1.4.0 已通过 `pending_marks` outbox 机制解决：FTS 写入成功但 vector_index_hash 标记失败时，下次启动自动 replay，不会触发重复 embedding。
+
+## 📜 版本历史
+
+- **v1.4.0** (2026-05): 数据架构与跨库一致性大修
+  - LanceDB chunks 表移除 `display_code/vector_text`，正文回查 `files.content`
+  - SemanticSplitter 偏移统一到 UTF-16 字符域
+  - schema_version 2 → 3，新增 `pending_marks` outbox + 三态迁移状态机
+  - 新增 `contextweaver migrate` CLI
+  - 跨进程 advisory lock 防止迁移竞争
+  - 109 个测试（含真实 LanceDB 端到端集成）
+- **v1.3.x**: 跨库写入事务性、scan 末尾自动 GC、files_fts 外部内容表
+- **v1.2.x**: 搜索管道优化、索引内存优化
+- **v1.1.x**: 智能 TopK 截断、Smart Cutoff
+- **v1.0.x**: 初始 release
+
 ## 📄 开源协议
 
 本项目采用 MIT 许可证。

package/dist/{SearchService-OS7CYHNJ.js → SearchService-WVD6THR3.js}

@@ -1,21 +1,32 @@
 import {
-  ChunkContentLoader,
+  createSearchConfigFingerprint
+} from "./chunk-IZ6IUHNN.js";
+import {
   bootstrap,
   getGraphExpander,
   getIndexer,
   scoreChunkTokenOverlap
-} from "./chunk-AB24E3Z7.js";
+} from "./chunk-XMZZZKG7.js";
+import "./chunk-LB42CZEB.js";
+import {
+  ChunkContentLoader
+} from "./chunk-XFIM2T6S.js";
 import {
   getVectorStore
-} from "./chunk-ZOMGPIU6.js";
+} from "./chunk-3BNHQV5W.js";
 import {
+  DEFAULT_CONFIG
+} from "./chunk-BFCIZ52F.js";
+import {
+  getIndexVersion,
+  incrementStat,
   initDb,
   isChunksFtsInitialized,
   isFtsInitialized,
   searchChunksFts,
   searchFilesFts,
   segmentQuery
-} from "./chunk-RGJSXUFS.js";
+} from "./chunk-PPLFJGO3.js";
 import {
   isDebugEnabled,
   logger
@@ -161,10 +172,8 @@ function sleep(ms) {
 
 // src/search/ContextPacker.ts
 var ContextPacker = class {
-  projectId;
   config;
-  constructor(projectId, config) {
-    this.projectId = projectId;
+  constructor(_projectId, config) {
     this.config = config;
   }
   /**
@@ -270,73 +279,63 @@ var ContextPacker = class {
   }
 };
 
-// src/search/config.ts
-var DEFAULT_CONFIG = {
-  // ── Recall (向量 + 词法召回) ──
-  vectorTopK: 80,
-  // Vector ANN candidates before dedup. Range: 40–200. Higher = better recall, more compute.
-  vectorTopM: 60,
-  // Vectors kept after dedup. Range: 30–100.
-  ftsTopKFiles: 20,
-  // Max files returned by FTS5 full-text search. Range: 10–50.
-  lexChunksPerFile: 2,
-  // Chunks to pull per FTS-matched file. Range: 1–5. Low keeps diversity across files.
-  lexTotalChunks: 40,
-  // Hard cap on total lexical chunks. Range: 20–80.
-  // ── RRF Fusion (向量 + 词法分数融合) ──
-  rrfK0: 20,
-  // RRF smoothing constant. Range: 10–60. Lower amplifies top ranks.
-  wVec: 0.6,
-  // Vector weight in fused score. Range: 0.3–0.8. Semantic relevance emphasis.
-  wLex: 0.4,
-  // Lexical weight in fused score. wVec + wLex should equal 1.0.
-  fusedTopM: 60,
-  // Candidates after fusion, fed into reranker. Range: 30–100.
-  // ── Rerank (精排) ──
-  rerankTopN: 10,
-  // Final top-N results after reranking. Range: 5–20.
-  maxRerankChars: 1e3,
-  // Max chars per chunk sent to reranker. Truncated beyond this. Range: 500–2000.
-  maxBreadcrumbChars: 250,
-  // Max chars for breadcrumb context in rerank input. Range: 100–500.
-  headRatio: 0.67,
-  // Ratio of head vs tail when truncating chunks. Range: 0.5–0.8.
-  // ── Expansion (上下文扩展: E1 邻居 / E2 面包屑 / E3 跨文件导入) ──
-  neighborHops: 2,
-  // E1: How many sibling chunks to expand in each direction. Range: 1–3.
-  breadcrumbExpandLimit: 3,
-  // E2: Max ancestor breadcrumbs (class/function scope). Range: 1–5.
-  importFilesPerSeed: 3,
-  // E3: Cross-file import files to resolve per seed chunk. Range: 0–5. Set to 3 to enable import-graph expansion for better cross-file context.
-  chunksPerImportFile: 3,
-  // E3: Chunks to pull from each resolved import file. Range: 1–5. Set to 3 for balanced coverage of imported symbols.
-  decayNeighbor: 0.8,
-  // Score decay per E1 hop. Range: 0.5–0.9. Higher = neighbors stay relevant longer.
-  decayBreadcrumb: 0.7,
-  // Score decay per E2 level. Range: 0.4–0.8.
-  decayImport: 0.6,
-  // Score decay for E3 import chunks. Range: 0.3–0.7. Lower than E1/E2 since cross-file is less certain.
-  decayDepth: 0.7,
-  // General depth decay multiplier. Range: 0.5–0.9.
-  // ── ContextPacker (上下文打包) ──
-  maxSegmentsPerFile: 3,
-  // Max non-contiguous segments per file in output. Range: 1–5. Prevents excessive fragmentation.
-  maxTotalChars: 48e3,
-  // Token budget expressed as chars (~12k tokens). Range: 20000–80000.
-  // ── Smart TopK (动态结果数量) ──
-  enableSmartTopK: true,
-  // Dynamically adjust result count based on score distribution.
-  smartTopScoreRatio: 0.5,
-  // Min score as ratio of top-1 score to remain included. Range: 0.3–0.7.
-  smartTopScoreDeltaAbs: 0.25,
-  // Max absolute score drop from top-1 before cutting off. Range: 0.1–0.4.
-  smartMinScore: 0.25,
-  // Hard floor: chunks below this score are always excluded. Range: 0.1–0.4.
-  smartMinK: 2,
-  // Minimum results to return regardless of scores. Range: 1–3.
-  smartMaxK: 8
-  // Maximum results when smart topK is active. Range: 5–15.
+// src/search/QueryCache.ts
+import crypto from "crypto";
+var MAX_CACHE_ENTRIES = 50;
+var LruCache = class {
+  constructor(maxSize) {
+    this.maxSize = maxSize;
+  }
+  entries = /* @__PURE__ */ new Map();
+  get(key) {
+    const value = this.entries.get(key);
+    if (value === void 0) return void 0;
+    this.entries.delete(key);
+    this.entries.set(key, value);
+    return value;
+  }
+  set(key, value) {
+    if (this.entries.has(key)) {
+      this.entries.delete(key);
+    }
+    this.entries.set(key, value);
+    if (this.entries.size > this.maxSize) {
+      const oldestKey = this.entries.keys().next().value;
+      if (oldestKey !== void 0) {
+        this.entries.delete(oldestKey);
+      }
+    }
+  }
 };
+var projectCaches = /* @__PURE__ */ new Map();
+function normalizeQuery(query) {
+  return query.trim().replace(/\s+/g, " ").toLowerCase();
+}
+function getProjectCache(projectId) {
+  let cache = projectCaches.get(projectId);
+  if (!cache) {
+    cache = new LruCache(MAX_CACHE_ENTRIES);
+    projectCaches.set(projectId, cache);
+  }
+  return cache;
+}
+function buildQueryCacheKey(input) {
+  const normalizedQuery = normalizeQuery(input.query);
+  return crypto.createHash("sha256").update(
+    JSON.stringify({
+      query: normalizedQuery,
+      projectId: input.projectId,
+      indexVersion: input.indexVersion,
+      configFingerprint: input.configFingerprint
+    })
+  ).digest("hex");
+}
+function getCachedContextPack(projectId, key) {
+  return getProjectCache(projectId).get(key);
+}
+function setCachedContextPack(projectId, key, pack) {
+  getProjectCache(projectId).set(key, pack);
+}
 
 // src/search/SearchService.ts
 var SearchService = class {
@@ -345,9 +344,11 @@ var SearchService = class {
   vectorStore = null;
   db = null;
   config;
+  configFingerprint;
   constructor(projectId, _projectPath, config) {
     this.projectId = projectId;
     this.config = { ...DEFAULT_CONFIG, ...config };
+    this.configFingerprint = createSearchConfigFingerprint(this.config);
   }
   async init() {
     const embeddingConfig = getEmbeddingConfig();
@@ -366,6 +367,18 @@ var SearchService = class {
    * 构建上下文包（用于问答/生成）
    */
   async buildContextPack(query) {
+    const db = this.db;
+    const cacheKey = buildQueryCacheKey({
+      query,
+      projectId: this.projectId,
+      indexVersion: getIndexVersion(db),
+      configFingerprint: this.configFingerprint
+    });
+    const cached = getCachedContextPack(this.projectId, cacheKey);
+    if (cached) {
+      this.recordSearchStats(db, { cacheHit: true });
+      return cached;
+    }
     const timingMs = {};
     let t0 = Date.now();
     const candidates = await this.hybridRetrieve(query);
@@ -385,7 +398,7 @@ var SearchService = class {
     const packer = new ContextPacker(this.projectId, this.config);
     const files = await packer.pack([...seeds, ...expanded], this.db);
     timingMs.pack = Date.now() - t0;
-    return {
+    const pack = {
       query,
       seeds,
       expanded,
@@ -396,6 +409,35 @@ var SearchService = class {
         timingMs
       }
     };
+    setCachedContextPack(this.projectId, cacheKey, pack);
+    this.recordSearchStats(db, { cacheHit: false, timingMs, seedCount: seeds.length });
+    return pack;
+  }
+  /**
+   * 记录搜索统计埋点（静默吞错，不影响搜索主流程）
+   *
+   * 多个计数器用事务包裹，保证一次查询要么全写要么全不写。
+   */
+  recordSearchStats(db, args) {
+    try {
+      const tx = db.transaction(() => {
+        incrementStat(db, "stats.search.total_queries");
+        if (args.cacheHit) {
+          incrementStat(db, "stats.search.cache_hits");
+          return;
+        }
+        incrementStat(db, "stats.search.compute_runs");
+        const t = args.timingMs ?? {};
+        incrementStat(db, "stats.search.sum_retrieve_ms", Math.round(t.retrieve ?? 0));
+        incrementStat(db, "stats.search.sum_rerank_ms", Math.round(t.rerank ?? 0));
+        incrementStat(db, "stats.search.sum_expand_ms", Math.round(t.expand ?? 0));
+        incrementStat(db, "stats.search.sum_pack_ms", Math.round(t.pack ?? 0));
+        incrementStat(db, "stats.search.sum_seed_count", args.seedCount ?? 0);
+      });
+      tx();
+    } catch (err) {
+      logger.debug({ error: err.message }, "\u641C\u7D22\u7EDF\u8BA1\u57CB\u70B9\u5931\u8D25");
+    }
   }
   // 召回方法
   /**

package/dist/{chunk-ZOMGPIU6.js → chunk-3BNHQV5W.js}

@@ -35,11 +35,9 @@ function sampleCheckDisplayCode(oldRows, getContent, options = {}) {
 var VectorStore = class {
   db = null;
   table = null;
-  projectId;
   dbPath;
   vectorDim;
   constructor(projectId, vectorDim = 1024, dbPathOverride) {
-    this.projectId = projectId;
     this.dbPath = dbPathOverride ?? path.join(BASE_DIR, projectId, "vectors.lance");
     this.vectorDim = vectorDim;
   }
@@ -185,9 +183,7 @@ var VectorStore = class {
         continue;
       }
       if (this.table && batch.length > 0) {
-        await this.deleteFilesByHash(
-          batch.map((f) => ({ path: f.path, hash: f.hash }))
-        );
+        await this.deleteFilesByHash(batch.map((f) => ({ path: f.path, hash: f.hash })));
       }
       if (!this.table) {
         await this.ensureTable(batchRecords);

package/dist/chunk-BFCIZ52F.js

@@ -0,0 +1,102 @@
+// src/search/config.ts
+var SEARCH_CONFIG_BOUNDS = {
+  vectorTopK: { min: 40, max: 200, integer: true },
+  vectorTopM: { min: 30, max: 100, integer: true },
+  ftsTopKFiles: { min: 10, max: 50, integer: true },
+  lexChunksPerFile: { min: 1, max: 5, integer: true },
+  lexTotalChunks: { min: 20, max: 80, integer: true },
+  rrfK0: { min: 10, max: 60, integer: true },
+  wVec: { min: 0, max: 1, integer: false },
+  wLex: { min: 0, max: 1, integer: false },
+  fusedTopM: { min: 30, max: 100, integer: true },
+  rerankTopN: { min: 5, max: 20, integer: true },
+  maxRerankChars: { min: 500, max: 2e3, integer: true },
+  maxBreadcrumbChars: { min: 100, max: 500, integer: true },
+  headRatio: { min: 0.5, max: 0.8, integer: false },
+  neighborHops: { min: 1, max: 3, integer: true },
+  breadcrumbExpandLimit: { min: 1, max: 5, integer: true },
+  importFilesPerSeed: { min: 0, max: 5, integer: true },
+  chunksPerImportFile: { min: 1, max: 5, integer: true },
+  decayNeighbor: { min: 0.5, max: 0.9, integer: false },
+  decayBreadcrumb: { min: 0.4, max: 0.8, integer: false },
+  decayImport: { min: 0.3, max: 0.7, integer: false },
+  decayDepth: { min: 0.5, max: 0.9, integer: false },
+  maxSegmentsPerFile: { min: 1, max: 5, integer: true },
+  maxTotalChars: { min: 2e4, max: 8e4, integer: true },
+  smartTopScoreRatio: { min: 0.3, max: 0.7, integer: false },
+  smartTopScoreDeltaAbs: { min: 0.1, max: 0.4, integer: false },
+  smartMinScore: { min: 0.1, max: 0.4, integer: false },
+  smartMinK: { min: 1, max: 3, integer: true },
+  smartMaxK: { min: 5, max: 15, integer: true }
+};
+var DEFAULT_CONFIG = {
+  // ── Recall (向量 + 词法召回) ──
+  vectorTopK: 80,
+  // Vector ANN candidates before dedup. Range: 40–200. Higher = better recall, more compute.
+  vectorTopM: 60,
+  // Vectors kept after dedup. Range: 30–100.
+  ftsTopKFiles: 20,
+  // Max files returned by FTS5 full-text search. Range: 10–50.
+  lexChunksPerFile: 2,
+  // Chunks to pull per FTS-matched file. Range: 1–5. Low keeps diversity across files.
+  lexTotalChunks: 40,
+  // Hard cap on total lexical chunks. Range: 20–80.
+  // ── RRF Fusion (向量 + 词法分数融合) ──
+  rrfK0: 20,
+  // RRF smoothing constant. Range: 10–60. Lower amplifies top ranks.
+  wVec: 0.6,
+  // Vector weight in fused score. Range: 0.3–0.8. Semantic relevance emphasis.
+  wLex: 0.4,
+  // Lexical weight in fused score. wVec + wLex should equal 1.0.
+  fusedTopM: 60,
+  // Candidates after fusion, fed into reranker. Range: 30–100.
+  // ── Rerank (精排) ──
+  rerankTopN: 10,
+  // Final top-N results after reranking. Range: 5–20.
+  maxRerankChars: 1e3,
+  // Max chars per chunk sent to reranker. Truncated beyond this. Range: 500–2000.
+  maxBreadcrumbChars: 250,
+  // Max chars for breadcrumb context in rerank input. Range: 100–500.
+  headRatio: 0.67,
+  // Ratio of head vs tail when truncating chunks. Range: 0.5–0.8.
+  // ── Expansion (上下文扩展: E1 邻居 / E2 面包屑 / E3 跨文件导入) ──
+  neighborHops: 2,
+  // E1: How many sibling chunks to expand in each direction. Range: 1–3.
+  breadcrumbExpandLimit: 3,
+  // E2: Max ancestor breadcrumbs (class/function scope). Range: 1–5.
+  importFilesPerSeed: 3,
+  // E3: Cross-file import files to resolve per seed chunk. Range: 0–5. Set to 3 to enable import-graph expansion for better cross-file context.
+  chunksPerImportFile: 3,
+  // E3: Chunks to pull from each resolved import file. Range: 1–5. Set to 3 for balanced coverage of imported symbols.
+  decayNeighbor: 0.8,
+  // Score decay per E1 hop. Range: 0.5–0.9. Higher = neighbors stay relevant longer.
+  decayBreadcrumb: 0.7,
+  // Score decay per E2 level. Range: 0.4–0.8.
+  decayImport: 0.6,
+  // Score decay for E3 import chunks. Range: 0.3–0.7. Lower than E1/E2 since cross-file is less certain.
+  decayDepth: 0.7,
+  // General depth decay multiplier. Range: 0.5–0.9.
+  // ── ContextPacker (上下文打包) ──
+  maxSegmentsPerFile: 3,
+  // Max non-contiguous segments per file in output. Range: 1–5. Prevents excessive fragmentation.
+  maxTotalChars: 48e3,
+  // Token budget expressed as chars (~12k tokens). Range: 20000–80000.
+  // ── Smart TopK (动态结果数量) ──
+  enableSmartTopK: true,
+  // Dynamically adjust result count based on score distribution.
+  smartTopScoreRatio: 0.5,
+  // Min score as ratio of top-1 score to remain included. Range: 0.3–0.7.
+  smartTopScoreDeltaAbs: 0.25,
+  // Max absolute score drop from top-1 before cutting off. Range: 0.1–0.4.
+  smartMinScore: 0.25,
+  // Hard floor: chunks below this score are always excluded. Range: 0.1–0.4.
+  smartMinK: 2,
+  // Minimum results to return regardless of scores. Range: 1–3.
+  smartMaxK: 8
+  // Maximum results when smart topK is active. Range: 5–15.
+};
+
+export {
+  SEARCH_CONFIG_BOUNDS,
+  DEFAULT_CONFIG
+};