ctxman 3.1.1 → 3.2.7

package/lib/rag/stores/SqliteVectorStore.js

@@ -0,0 +1,595 @@
+import fs from 'fs';
+import path from 'path';
+import crypto from 'crypto';
+import { createRequire } from 'module';
+import { BaseVectorStore } from './BaseVectorStore.js';
+
+const require = createRequire(import.meta.url);
+let cachedDatabaseSync = null;
+
+/**
+ * Lazily loads `DatabaseSync` from `node:sqlite` while suppressing Node's
+ * one-time `ExperimentalWarning: SQLite is an experimental feature` on stderr.
+ */
+function getDatabaseSync() {
+  if (cachedDatabaseSync) return cachedDatabaseSync;
+  const origEmitWarning = process.emitWarning;
+  process.emitWarning = function (warning, ...args) {
+    const msg = typeof warning === 'string' ? warning : warning?.message || '';
+    if (/SQLite is an experimental feature/i.test(msg)) return;
+    return origEmitWarning.call(this, warning, ...args);
+  };
+  try {
+    ({ DatabaseSync: cachedDatabaseSync } = require('node:sqlite'));
+    return cachedDatabaseSync;
+  } finally {
+    process.emitWarning = origEmitWarning;
+  }
+}
+
+const toPosix = (p) =>
+  String(p ?? '')
+    .split(path.sep)
+    .join('/');
+
+/** Byte-order comparison, identical across platforms and locales */
+function comparePaths(a, b) {
+  return a < b ? -1 : a > b ? 1 : 0;
+}
+
+/**
+ * Zero-native-binary vector & knowledge-graph store using Node 22+ built-in node:sqlite (`DatabaseSync`).
+ * Stores document chunks, content SHA-256 hashes, metadata, float32 embedding vectors,
+ * and incremental repository knowledge-graph nodes & edges in `<storePath>/index.sqlite`.
+ * Supports incremental upserts (skipping unchanged content), document removal,
+ * graph delta synchronization, and exact cosine similarity vector search.
+ */
+export class SqliteVectorStore extends BaseVectorStore {
+  /**
+   * @param {import('../EmbeddingProvider').EmbeddingProvider} embeddingProvider
+   * @param {string} [storagePath='.ctxman/self-knowledge'] - Directory or .sqlite file path
+   */
+  constructor(embeddingProvider, storagePath = '.ctxman/self-knowledge') {
+    super(embeddingProvider);
+    this.storagePath = storagePath;
+    this.dbPath = storagePath.endsWith('.sqlite')
+      ? storagePath
+      : path.join(storagePath, 'index.sqlite');
+    this.db = null;
+    this.initialized = false;
+  }
+
+  /**
+   * Initialize directory, SQLite database, documents table, and knowledge-graph tables.
+   */
+  async init() {
+    if (this.initialized) return;
+
+    const dir = path.dirname(this.dbPath);
+    if (!fs.existsSync(dir)) {
+      fs.mkdirSync(dir, { recursive: true });
+    }
+
+    const DatabaseSync = getDatabaseSync();
+    this.db = new DatabaseSync(this.dbPath);
+
+    // Create tables for documents (vectors) and knowledge graph (nodes + edges)
+    this.db.exec(`
+      CREATE TABLE IF NOT EXISTS documents (
+        id TEXT PRIMARY KEY,
+        file_path TEXT,
+        content_hash TEXT NOT NULL,
+        text TEXT NOT NULL,
+        metadata TEXT,
+        vector BLOB,
+        updated_at INTEGER
+      );
+      CREATE INDEX IF NOT EXISTS idx_documents_file_path ON documents(file_path);
+      CREATE INDEX IF NOT EXISTS idx_documents_content_hash ON documents(content_hash);
+
+      CREATE TABLE IF NOT EXISTS graph_nodes (
+        path TEXT PRIMARY KEY,
+        language TEXT NOT NULL,
+        bytes INTEGER NOT NULL DEFAULT 0,
+        content_hash TEXT,
+        methods TEXT NOT NULL DEFAULT '[]',
+        declarations TEXT NOT NULL DEFAULT '[]',
+        imports TEXT NOT NULL DEFAULT '[]',
+        symbol_refs REAL NOT NULL DEFAULT 0,
+        unresolved_imports INTEGER NOT NULL DEFAULT 0,
+        updated_at INTEGER
+      );
+
+      CREATE TABLE IF NOT EXISTS graph_edges (
+        from_path TEXT NOT NULL,
+        to_path TEXT NOT NULL,
+        weight REAL NOT NULL DEFAULT 1,
+        updated_at INTEGER,
+        PRIMARY KEY (from_path, to_path)
+      );
+      CREATE INDEX IF NOT EXISTS idx_graph_edges_from ON graph_edges(from_path);
+      CREATE INDEX IF NOT EXISTS idx_graph_edges_to ON graph_edges(to_path);
+    `);
+
+    this.stmtGetById = this.db.prepare('SELECT id, content_hash FROM documents WHERE id = ?');
+    this.stmtGetByPath = this.db.prepare(
+      'SELECT id, content_hash FROM documents WHERE file_path = ?'
+    );
+    this.stmtUpsert = this.db.prepare(`
+      INSERT INTO documents (id, file_path, content_hash, text, metadata, vector, updated_at)
+      VALUES (?, ?, ?, ?, ?, ?, ?)
+      ON CONFLICT(id) DO UPDATE SET
+        file_path = excluded.file_path,
+        content_hash = excluded.content_hash,
+        text = excluded.text,
+        metadata = excluded.metadata,
+        vector = excluded.vector,
+        updated_at = excluded.updated_at
+    `);
+    this.stmtDeleteById = this.db.prepare('DELETE FROM documents WHERE id = ?');
+    this.stmtAllMeta = this.db.prepare('SELECT id, file_path, metadata FROM documents');
+    this.stmtAllVectors = this.db.prepare(
+      'SELECT id, file_path, text, metadata, vector FROM documents'
+    );
+    this.stmtCount = this.db.prepare('SELECT COUNT(*) AS count FROM documents');
+
+    // Graph prepared statements
+    this.stmtUpsertNode = this.db.prepare(`
+      INSERT INTO graph_nodes (
+        path, language, bytes, content_hash, methods, declarations, imports,
+        symbol_refs, unresolved_imports, updated_at
+      )
+      VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
+      ON CONFLICT(path) DO UPDATE SET
+        language = excluded.language,
+        bytes = excluded.bytes,
+        content_hash = excluded.content_hash,
+        methods = excluded.methods,
+        declarations = excluded.declarations,
+        imports = excluded.imports,
+        symbol_refs = excluded.symbol_refs,
+        unresolved_imports = excluded.unresolved_imports,
+        updated_at = excluded.updated_at
+    `);
+    this.stmtDeleteNode = this.db.prepare('DELETE FROM graph_nodes WHERE path = ?');
+    this.stmtDeleteOutEdges = this.db.prepare('DELETE FROM graph_edges WHERE from_path = ?');
+    this.stmtDeleteIncidentEdges = this.db.prepare(
+      'DELETE FROM graph_edges WHERE from_path = ? OR to_path = ?'
+    );
+    this.stmtUpsertEdge = this.db.prepare(`
+      INSERT INTO graph_edges (from_path, to_path, weight, updated_at)
+      VALUES (?, ?, ?, ?)
+      ON CONFLICT(from_path, to_path) DO UPDATE SET
+        weight = excluded.weight,
+        updated_at = excluded.updated_at
+    `);
+    this.stmtAllNodes = this.db.prepare(`
+      SELECT path, language, bytes, content_hash, methods, declarations, imports,
+             symbol_refs, unresolved_imports, updated_at
+      FROM graph_nodes
+      ORDER BY path ASC
+    `);
+    this.stmtAllEdges = this.db.prepare(`
+      SELECT from_path, to_path, weight
+      FROM graph_edges
+      ORDER BY from_path ASC, to_path ASC
+    `);
+    this.stmtNodeCount = this.db.prepare('SELECT COUNT(*) AS count FROM graph_nodes');
+    this.stmtEdgeCount = this.db.prepare('SELECT COUNT(*) AS count FROM graph_edges');
+
+    this.initialized = true;
+  }
+
+  /**
+   * Add a single document to the store.
+   * @param {string} text
+   * @param {Object} [metadata={}]
+   * @returns {Promise<{upserted: number, skipped: number}>}
+   */
+  async addDocument(text, metadata = {}) {
+    return this.addDocuments([{ text, metadata }]);
+  }
+
+  /**
+   * Incremental upsert of documents.
+   * Skips re-embedding and rewriting if content SHA-256 is unchanged.
+   * @param {Array<{text: string, id?: string, metadata?: Object}>} docs
+   * @returns {Promise<{upserted: number, skipped: number}>}
+   */
+  async addDocuments(docs) {
+    if (!docs || docs.length === 0) return { upserted: 0, skipped: 0 };
+    if (!this.initialized) await this.init();
+
+    let upserted = 0;
+    let skipped = 0;
+
+    for (const doc of docs) {
+      const text = String(doc.text ?? '');
+      const contentHash = crypto.createHash('sha256').update(text).digest('hex');
+
+      const meta = doc.metadata ?? {};
+      const filePath =
+        meta.path ?? (typeof doc.id === 'string' && !doc.id.includes(':') ? doc.id : null);
+      const id = doc.id ?? meta.sourceId ?? (filePath ? filePath : crypto.randomUUID());
+
+      // Check if unchanged by id or filePath
+      const existing =
+        this.stmtGetById.get(id) ?? (filePath ? this.stmtGetByPath.get(filePath) : null);
+      if (existing && existing.content_hash === contentHash) {
+        // Content hash matches; skip re-embedding
+        skipped++;
+        continue;
+      }
+
+      const targetId = existing ? existing.id : id;
+
+      // Compute vector embedding
+      const vector = await this.embeddingProvider.embed(text);
+      const float32 = new Float32Array(vector);
+      const blob = Buffer.from(float32.buffer, float32.byteOffset, float32.byteLength);
+
+      this.stmtUpsert.run(
+        targetId,
+        filePath,
+        contentHash,
+        text,
+        JSON.stringify(meta),
+        blob,
+        Date.now()
+      );
+      upserted++;
+    }
+
+    return { upserted, skipped };
+  }
+
+  /**
+   * Remove document(s) matching a document ID, relative path, or absolute file path.
+   * @param {string} idOrFilePath
+   * @returns {Promise<number>} Number of removed documents
+   */
+  async removeDocument(idOrFilePath) {
+    if (!idOrFilePath) return 0;
+    if (!this.initialized) await this.init();
+
+    const target = String(idOrFilePath);
+    const posixTarget = toPosix(target);
+    const suffixTarget = posixTarget.startsWith('/') ? posixTarget : `/${posixTarget}`;
+
+    const rows = this.stmtAllMeta.all();
+    const idsToDelete = new Set();
+
+    for (const row of rows) {
+      if (row.id === target || row.id === posixTarget) {
+        idsToDelete.add(row.id);
+        continue;
+      }
+      if (row.file_path) {
+        const posixFilePath = toPosix(row.file_path);
+        if (
+          row.file_path === target ||
+          posixFilePath === posixTarget ||
+          posixFilePath.endsWith(suffixTarget)
+        ) {
+          idsToDelete.add(row.id);
+          continue;
+        }
+      }
+      if (row.metadata) {
+        try {
+          const meta = JSON.parse(row.metadata);
+          const metaPath = meta.path ? toPosix(meta.path) : null;
+          const metaRel = meta.relativePath ? toPosix(meta.relativePath) : null;
+          if (
+            metaPath === posixTarget ||
+            (metaPath && metaPath.endsWith(suffixTarget)) ||
+            metaRel === posixTarget
+          ) {
+            idsToDelete.add(row.id);
+          }
+        } catch {
+          // ignore malformed metadata
+        }
+      }
+    }
+
+    let removed = 0;
+    for (const id of idsToDelete) {
+      const res = this.stmtDeleteById.run(id);
+      removed += Number(res?.changes ?? 1);
+    }
+    return removed;
+  }
+
+  /**
+   * Upsert a single repository knowledge-graph node.
+   * @param {Object} node
+   */
+  async upsertGraphNode(node) {
+    if (!node || !node.path) return;
+    if (!this.initialized) await this.init();
+
+    const relPath = toPosix(node.path);
+    this.stmtUpsertNode.run(
+      relPath,
+      String(node.language ?? 'unknown'),
+      Number(node.bytes ?? 0),
+      node.contentHash ?? null,
+      JSON.stringify(node.methods ?? []),
+      JSON.stringify(node.declarations ?? []),
+      JSON.stringify(node.imports ?? []),
+      Number(node.symbolRefs ?? 0),
+      Number(node.unresolvedImports ?? 0),
+      Date.now()
+    );
+  }
+
+  /**
+   * Replace all outgoing import edges for a single file node.
+   * @param {string} fromPath
+   * @param {Array<{to: string, weight?: number}>} edges
+   */
+  async replaceGraphEdges(fromPath, edges = []) {
+    if (!fromPath) return;
+    if (!this.initialized) await this.init();
+
+    const posixFrom = toPosix(fromPath);
+    const now = Date.now();
+    this.db.exec('BEGIN IMMEDIATE');
+    try {
+      this.stmtDeleteOutEdges.run(posixFrom);
+      for (const edge of edges) {
+        if (!edge || !edge.to) continue;
+        this.stmtUpsertEdge.run(posixFrom, toPosix(edge.to), Number(edge.weight ?? 1), now);
+      }
+      this.db.exec('COMMIT');
+    } catch (err) {
+      this.db.exec('ROLLBACK');
+      throw err;
+    }
+  }
+
+  /**
+   * Remove a node and all incident (incoming + outgoing) edges from the knowledge graph.
+   * @param {string} filePath - Project-relative path
+   * @returns {Promise<{removedNode: number, removedEdges: number}>}
+   */
+  async removeGraphNode(filePath) {
+    if (!filePath) return { removedNode: 0, removedEdges: 0 };
+    if (!this.initialized) await this.init();
+
+    const posixPath = toPosix(filePath);
+    this.db.exec('BEGIN IMMEDIATE');
+    try {
+      const edgeRes = this.stmtDeleteIncidentEdges.run(posixPath, posixPath);
+      const nodeRes = this.stmtDeleteNode.run(posixPath);
+      this.db.exec('COMMIT');
+      return {
+        removedNode: Number(nodeRes?.changes ?? 0),
+        removedEdges: Number(edgeRes?.changes ?? 0),
+      };
+    } catch (err) {
+      this.db.exec('ROLLBACK');
+      throw err;
+    }
+  }
+
+  /**
+   * Synchronize a full or incremental set of nodes and edges in a single SQLite transaction.
+   * @param {{nodes: Array<Object>, edges: Array<{from: string, to: string, weight?: number}>, replaceAll?: boolean}} graph
+   */
+  async syncGraph(graph) {
+    if (!graph) return;
+    if (!this.initialized) await this.init();
+
+    const now = Date.now();
+    this.db.exec('BEGIN IMMEDIATE');
+    try {
+      if (graph.replaceAll) {
+        this.db.exec('DELETE FROM graph_nodes; DELETE FROM graph_edges;');
+      }
+      for (const node of graph.nodes ?? []) {
+        const relPath = toPosix(node.path);
+        this.stmtUpsertNode.run(
+          relPath,
+          String(node.language ?? 'unknown'),
+          Number(node.bytes ?? 0),
+          node.contentHash ?? null,
+          JSON.stringify(node.methods ?? []),
+          JSON.stringify(node.declarations ?? []),
+          JSON.stringify(node.imports ?? []),
+          Number(node.symbolRefs ?? 0),
+          Number(node.unresolvedImports ?? 0),
+          now
+        );
+      }
+      if (Array.isArray(graph.edges)) {
+        this.db.exec('DELETE FROM graph_edges;');
+        for (const edge of graph.edges) {
+          if (!edge || !edge.from || !edge.to) continue;
+          this.stmtUpsertEdge.run(
+            toPosix(edge.from),
+            toPosix(edge.to),
+            Number(edge.weight ?? 1),
+            now
+          );
+        }
+      }
+      this.db.exec('COMMIT');
+    } catch (err) {
+      this.db.exec('ROLLBACK');
+      throw err;
+    }
+  }
+
+  /**
+   * Read the persisted knowledge graph (`nodes` and `edges`) from SQLite.
+   * @returns {Promise<{nodes: Array<Object>, edges: Array<{from: string, to: string, weight: number}>}>}
+   */
+  async getGraph() {
+    if (!this.initialized) await this.init();
+
+    const rawNodes = this.stmtAllNodes.all();
+    const rawEdges = this.stmtAllEdges.all();
+
+    const parseJson = (str, fallback) => {
+      if (!str) return fallback;
+      try {
+        return JSON.parse(str);
+      } catch {
+        return fallback;
+      }
+    };
+
+    const nodes = rawNodes
+      .map((row) => ({
+        path: row.path,
+        language: row.language,
+        bytes: Number(row.bytes ?? 0),
+        contentHash: row.content_hash ?? null,
+        methods: parseJson(row.methods, []),
+        declarations: parseJson(row.declarations, []),
+        imports: parseJson(row.imports, []),
+        symbolRefs: Number(row.symbol_refs ?? 0),
+        unresolvedImports: Number(row.unresolved_imports ?? 0),
+        updatedAt: row.updated_at ? Number(row.updated_at) : null,
+      }))
+      .sort((a, b) => comparePaths(a.path, b.path));
+
+    const edges = rawEdges
+      .map((row) => ({
+        from: row.from_path,
+        to: row.to_path,
+        weight: Number(row.weight ?? 1),
+      }))
+      .sort((a, b) => comparePaths(a.from, b.from) || comparePaths(a.to, b.to));
+
+    return { nodes, edges };
+  }
+
+  /**
+   * Return node and edge counts in the persisted knowledge graph.
+   * @returns {{nodes: number, edges: number}}
+   */
+  graphCounts() {
+    if (!this.initialized || !this.db) return { nodes: 0, edges: 0 };
+    const nRow = this.stmtNodeCount.get();
+    const eRow = this.stmtEdgeCount.get();
+    return {
+      nodes: nRow ? Number(nRow.count) : 0,
+      edges: eRow ? Number(eRow.count) : 0,
+    };
+  }
+
+  /**
+   * Exact cosine similarity search over stored documents.
+   * @param {string} queryText
+   * @param {number} [limit=5]
+   * @returns {Promise<Array<{id?: string, text: string, score: number, metadata: Object}>>}
+   */
+  async search(queryText, limit = 5) {
+    if (!this.initialized) await this.init();
+
+    const countRow = this.stmtCount.get();
+    if (!countRow || Number(countRow.count) === 0) return [];
+
+    const queryVector = await this.embeddingProvider.embed(queryText);
+    const qLen = queryVector.length;
+
+    let queryNormSq = 0;
+    for (let i = 0; i < qLen; i++) {
+      queryNormSq += queryVector[i] * queryVector[i];
+    }
+    const queryNorm = Math.sqrt(queryNormSq);
+    if (queryNorm === 0) return [];
+
+    const rows = this.stmtAllVectors.all();
+    const scored = [];
+
+    for (const row of rows) {
+      if (!row.vector) continue;
+      const buf = row.vector;
+      const docVector = new Float32Array(buf.buffer, buf.byteOffset, buf.byteLength / 4);
+
+      let dot = 0;
+      let docNormSq = 0;
+      const dim = Math.min(qLen, docVector.length);
+
+      for (let i = 0; i < dim; i++) {
+        const qVal = queryVector[i];
+        const dVal = docVector[i];
+        dot += qVal * dVal;
+        docNormSq += dVal * dVal;
+      }
+
+      const docNorm = Math.sqrt(docNormSq);
+      const similarity = docNorm === 0 ? 0 : dot / (queryNorm * docNorm);
+
+      let parsedMeta = {};
+      if (row.metadata) {
+        try {
+          parsedMeta = JSON.parse(row.metadata);
+        } catch {
+          parsedMeta = {};
+        }
+      }
+
+      scored.push({
+        id: row.id,
+        text: row.text,
+        score: similarity,
+        metadata: parsedMeta,
+      });
+    }
+
+    // Sort descending by similarity score
+    scored.sort((a, b) => b.score - a.score);
+
+    return scored.slice(0, limit);
+  }
+
+  /**
+   * Delete documents and graph tables and reinitialize.
+   */
+  async deleteTable() {
+    if (!this.initialized) await this.init();
+    try {
+      this.db.exec(`
+        DROP TABLE IF EXISTS documents;
+        DROP TABLE IF EXISTS graph_nodes;
+        DROP TABLE IF EXISTS graph_edges;
+      `);
+      this.initialized = false;
+      await this.init();
+    } catch (_e) {
+      // Ignore if table does not exist
+    }
+  }
+
+  /**
+   * Total number of indexed documents.
+   * @returns {number}
+   */
+  count() {
+    if (!this.initialized || !this.db) return 0;
+    const row = this.stmtCount.get();
+    return row ? Number(row.count) : 0;
+  }
+
+  /**
+   * Close the database connection.
+   */
+  close() {
+    if (this.db) {
+      try {
+        this.db.close();
+      } catch (_e) {
+        // ignore
+      }
+      this.db = null;
+      this.initialized = false;
+    }
+  }
+}
+
+export default SqliteVectorStore;