kirograph 0.12.1

package/dist/vectors/index.js

@@ -0,0 +1,480 @@
+"use strict";
+var __create = Object.create;
+var __defProp = Object.defineProperty;
+var __getOwnPropDesc = Object.getOwnPropertyDescriptor;
+var __getOwnPropNames = Object.getOwnPropertyNames;
+var __getProtoOf = Object.getPrototypeOf;
+var __hasOwnProp = Object.prototype.hasOwnProperty;
+var __export = (target, all) => {
+  for (var name in all)
+    __defProp(target, name, { get: all[name], enumerable: true });
+};
+var __copyProps = (to, from, except, desc) => {
+  if (from && typeof from === "object" || typeof from === "function") {
+    for (let key of __getOwnPropNames(from))
+      if (!__hasOwnProp.call(to, key) && key !== except)
+        __defProp(to, key, { get: () => from[key], enumerable: !(desc = __getOwnPropDesc(from, key)) || desc.enumerable });
+  }
+  return to;
+};
+var __toESM = (mod, isNodeMode, target) => (target = mod != null ? __create(__getProtoOf(mod)) : {}, __copyProps(
+  // If the importer is in node compatibility mode or this is not an ESM
+  // file that has been converted to a CommonJS file using a Babel-
+  // compatible transform (i.e. "__esModule" has not been set), then set
+  // "default" to the CommonJS "module.exports" for node compatibility.
+  isNodeMode || !mod || !mod.__esModule ? __defProp(target, "default", { value: mod, enumerable: true }) : target,
+  mod
+));
+var __toCommonJS = (mod) => __copyProps(__defProp({}, "__esModule", { value: true }), mod);
+var vectors_exports = {};
+__export(vectors_exports, {
+  VectorManager: () => VectorManager
+});
+module.exports = __toCommonJS(vectors_exports);
+var path = __toESM(require("path"));
+var fs = __toESM(require("fs"));
+var import_os = require("os");
+var import_errors = require("../errors");
+var import_vec_index = require("./vec-index");
+var import_orama_index = require("./orama-index");
+var import_pglite_index = require("./pglite-index");
+var import_lancedb_index = require("./lancedb-index");
+var import_qdrant_index = require("./qdrant-index");
+var import_typesense_index = require("./typesense-index");
+const DEFAULT_MODEL = "nomic-ai/nomic-embed-text-v1.5";
+const DEFAULT_EMBEDDING_DIM = 768;
+const GLOBAL_MODELS_DIR = path.join((0, import_os.homedir)(), ".kirograph", "models");
+const BATCH_SIZE = 32;
+const MIN_BATCH_SIZE = 4;
+const MAX_TOKEN_CHARS = 2e3;
+const EMBEDDABLE_KINDS = /* @__PURE__ */ new Set([
+  "function",
+  "method",
+  "class",
+  "interface",
+  "type_alias",
+  "component",
+  "module"
+]);
+let transformers = null;
+async function getTransformers() {
+  if (!transformers) {
+    transformers = await import("@huggingface/transformers");
+  }
+  return transformers;
+}
+function nodeToText(node) {
+  const parts = [`${node.kind}: ${node.name}`];
+  if (node.qualifiedName && node.qualifiedName !== node.name) {
+    parts.push(`path: ${node.qualifiedName}`);
+  }
+  parts.push(`file: ${node.filePath}`);
+  if (node.signature) parts.push(`signature: ${node.signature}`);
+  if (node.docstring) parts.push(`documentation: ${node.docstring}`);
+  const text = parts.join("\n");
+  if (text.length > MAX_TOKEN_CHARS) {
+    (0, import_errors.logDebug)(`nodeToText: truncated ${node.id} from ${text.length} to ${MAX_TOKEN_CHARS} chars`);
+    return text.slice(0, MAX_TOKEN_CHARS);
+  }
+  return text;
+}
+function cosine(a, b) {
+  let dot = 0, na = 0, nb = 0;
+  for (let i = 0; i < a.length; i++) {
+    dot += a[i] * b[i];
+    na += a[i] * a[i];
+    nb += b[i] * b[i];
+  }
+  const denom = Math.sqrt(na) * Math.sqrt(nb);
+  return denom === 0 ? 0 : dot / denom;
+}
+function toFloat32Array(data) {
+  if (data instanceof Float32Array) return data;
+  if (Array.isArray(data)) return new Float32Array(data);
+  if (data && typeof data === "object" && "length" in data) {
+    return Float32Array.from(Array.from(data));
+  }
+  throw new Error("Unsupported embedding data format");
+}
+class VectorManager {
+  constructor(db, config, projectRoot) {
+    this.db = db;
+    this.config = config;
+    this.projectRoot = projectRoot;
+    this.pipeline = null;
+    this._initialized = false;
+    this.vecIndex = null;
+    this.oramaIndex = null;
+    this.pgliteIndex = null;
+    this.lancedbIndex = null;
+    this.qdrantIndex = null;
+    this.typesenseIndex = null;
+    this._engineFallback = null;
+  }
+  isInitialized() {
+    return this.config.enableEmbeddings === true && this._initialized;
+  }
+  /** Non-null when the configured engine failed to load and cosine is being used instead. */
+  getEngineFallback() {
+    return this._engineFallback;
+  }
+  /**
+   * Load the embedding model. No-op when embeddings are disabled.
+   * Fails silently so callers can continue without semantic search.
+   * When config.useVecIndex is true, also initializes the sqlite-vec ANN index.
+   */
+  async initialize(onProgress) {
+    if (!this.config.enableEmbeddings) {
+      (0, import_errors.logDebug)("VectorManager: embeddings disabled");
+      return;
+    }
+    if (this._initialized) return;
+    const rawModel = this.config.embeddingModel || DEFAULT_MODEL;
+    const modelId = rawModel.includes("/") ? rawModel : DEFAULT_MODEL;
+    if (rawModel !== modelId) {
+      (0, import_errors.logWarn)(`VectorManager: invalid embeddingModel "${rawModel}", using default instead`, { default: DEFAULT_MODEL });
+    }
+    const cacheDir = GLOBAL_MODELS_DIR;
+    try {
+      const { pipeline, env } = await getTransformers();
+      env.cacheDir = cacheDir;
+      if (!fs.existsSync(cacheDir)) {
+        fs.mkdirSync(cacheDir, { recursive: true });
+      }
+      const cached = fs.existsSync(path.join(cacheDir, modelId));
+      if (cached) env.allowRemoteModels = false;
+      let lastFile = "";
+      this.pipeline = await pipeline("feature-extraction", modelId, {
+        dtype: "auto",
+        ...!cached && onProgress ? {
+          progress_callback: (p) => {
+            if (p.status === "progress" && p.file) {
+              lastFile = p.file;
+              onProgress(p.file, p.loaded ?? 0, p.total ?? 0, false);
+            } else if (p.status === "done" && lastFile) {
+              onProgress(lastFile, 1, 1, true);
+              lastFile = "";
+            }
+          }
+        } : {}
+      });
+      try {
+        const testOut = await this.pipeline("test", { pooling: "mean", normalize: true });
+        const actualDim = testOut.dims?.[1] ?? testOut.data?.length;
+        const configuredDim = this.config.embeddingDim ?? DEFAULT_EMBEDDING_DIM;
+        if (actualDim && actualDim !== configuredDim) {
+          (0, import_errors.logWarn)(
+            `VectorManager: embeddingDim mismatch \u2014 config says ${configuredDim} but model outputs ${actualDim}. Using actual dim.`,
+            { model: modelId, configuredDim, actualDim }
+          );
+          this.config.embeddingDim = actualDim;
+        }
+      } catch {
+      }
+      this._initialized = true;
+      (0, import_errors.logDebug)("VectorManager: model loaded", { modelId, embeddingDim: this.config.embeddingDim ?? DEFAULT_EMBEDDING_DIM });
+    } catch (err) {
+      (0, import_errors.logError)("VectorManager: failed to load embedding model", {
+        model: modelId,
+        error: String(err)
+      });
+      this._initialized = false;
+      return;
+    }
+    const engine = this.config.semanticEngine ?? (this.config.useVecIndex ? "sqlite-vec" : "cosine");
+    if (this.projectRoot && engine !== "cosine") {
+      const kirographDir = path.join(this.projectRoot, ".kirograph");
+      const embeddingDim = this.config.embeddingDim ?? DEFAULT_EMBEDDING_DIM;
+      if (engine === "sqlite-vec") {
+        this.vecIndex = new import_vec_index.VecIndex(kirographDir, embeddingDim);
+        await this.vecIndex.initialize();
+        if (this.vecIndex.isAvailable()) {
+          (0, import_errors.logDebug)("VectorManager: sqlite-vec ANN index ready");
+        } else {
+          this._engineFallback = "sqlite-vec unavailable \u2014 run: npm install better-sqlite3 sqlite-vec";
+          (0, import_errors.logDebug)("VectorManager: sqlite-vec unavailable, falling back to in-process cosine");
+        }
+      } else if (engine === "orama") {
+        this.oramaIndex = new import_orama_index.OramaIndex(kirographDir, embeddingDim);
+        await this.oramaIndex.initialize();
+        if (this.oramaIndex.isAvailable()) {
+          (0, import_errors.logDebug)("VectorManager: Orama hybrid index ready");
+        } else {
+          this._engineFallback = "orama unavailable \u2014 run: npm install @orama/orama @orama/plugin-data-persistence";
+          (0, import_errors.logDebug)("VectorManager: Orama unavailable, falling back to in-process cosine");
+        }
+      } else if (engine === "pglite") {
+        this.pgliteIndex = new import_pglite_index.PGliteIndex(kirographDir, embeddingDim);
+        await this.pgliteIndex.initialize();
+        if (this.pgliteIndex.isAvailable()) {
+          (0, import_errors.logDebug)("VectorManager: PGlite+pgvector hybrid index ready");
+        } else {
+          this._engineFallback = "pglite unavailable \u2014 run: npm install @electric-sql/pglite";
+          (0, import_errors.logDebug)("VectorManager: PGlite unavailable, falling back to in-process cosine");
+        }
+      } else if (engine === "lancedb") {
+        this.lancedbIndex = new import_lancedb_index.LanceDBIndex(kirographDir, embeddingDim);
+        await this.lancedbIndex.initialize();
+        if (this.lancedbIndex.isAvailable()) {
+          (0, import_errors.logDebug)("VectorManager: LanceDB ANN index ready");
+        } else {
+          this._engineFallback = "lancedb unavailable \u2014 run: npm install @lancedb/lancedb";
+          (0, import_errors.logDebug)("VectorManager: LanceDB unavailable, falling back to in-process cosine");
+        }
+      } else if (engine === "qdrant") {
+        this.qdrantIndex = new import_qdrant_index.QdrantIndex(kirographDir, embeddingDim);
+        await this.qdrantIndex.initialize();
+        if (this.qdrantIndex.isAvailable()) {
+          (0, import_errors.logDebug)("VectorManager: Qdrant ANN index ready");
+        } else {
+          this._engineFallback = "qdrant unavailable \u2014 run: npm install qdrant-local";
+          (0, import_errors.logDebug)("VectorManager: Qdrant unavailable, falling back to in-process cosine");
+        }
+      } else if (engine === "typesense") {
+        this.typesenseIndex = new import_typesense_index.TypesenseIndex(kirographDir, embeddingDim);
+        await this.typesenseIndex.initialize();
+        if (this.typesenseIndex.isAvailable()) {
+          (0, import_errors.logDebug)("VectorManager: Typesense ANN index ready");
+        } else {
+          const reason = this.typesenseIndex.getFailReason();
+          this._engineFallback = reason ? `typesense initialization failed: ${reason}` : "typesense unavailable \u2014 run: npm install typesense";
+          (0, import_errors.logDebug)("VectorManager: Typesense unavailable, falling back to in-process cosine");
+        }
+      }
+    }
+  }
+  /**
+   * Embed a single node and persist to the vectors table.
+   * Skips silently when disabled; logs on failure without throwing.
+   */
+  async embedNode(node) {
+    if (!this.config.enableEmbeddings) return;
+    if (!this._initialized || !this.pipeline) {
+      (0, import_errors.logError)("Embedding model unavailable", { model: this.config.embeddingModel });
+      return;
+    }
+    if (!EMBEDDABLE_KINDS.has(node.kind)) return;
+    try {
+      const text = `search_document: ${nodeToText(node)}`;
+      const output = await this.pipeline(text, { pooling: "mean", normalize: true });
+      const embedding = toFloat32Array(output.data);
+      if (!this.vecIndex?.isAvailable() && !this.oramaIndex?.isAvailable() && !this.pgliteIndex?.isAvailable() && !this.lancedbIndex?.isAvailable() && !this.qdrantIndex?.isAvailable() && !this.typesenseIndex?.isAvailable()) {
+        this.db.storeEmbedding(node.id, embedding, this.config.embeddingModel || DEFAULT_MODEL);
+      }
+      this.vecIndex?.upsert(node.id, embedding);
+      if (this.oramaIndex?.isAvailable()) await this.oramaIndex.upsert(node, embedding);
+      if (this.pgliteIndex?.isAvailable()) await this.pgliteIndex.upsert(node, embedding);
+      if (this.lancedbIndex?.isAvailable()) await this.lancedbIndex.upsert(node, embedding);
+      if (this.qdrantIndex?.isAvailable()) await this.qdrantIndex.upsert(node, embedding);
+      if (this.typesenseIndex?.isAvailable()) await this.typesenseIndex.upsert(node, embedding);
+    } catch (err) {
+      (0, import_errors.logWarn)("Failed to embed node", { nodeId: node.id, error: String(err) });
+    }
+  }
+  /** Number of entries currently in the active ANN/hybrid index (0 when not in use). */
+  async vecIndexCount() {
+    if (this.vecIndex?.isAvailable()) return this.vecIndex.count();
+    if (this.oramaIndex?.isAvailable()) return this.oramaIndex.count();
+    if (this.pgliteIndex?.isAvailable()) return this.pgliteIndex.count();
+    if (this.lancedbIndex?.isAvailable()) return this.lancedbIndex.count();
+    if (this.qdrantIndex?.isAvailable()) return this.qdrantIndex.count();
+    if (this.typesenseIndex?.isAvailable()) return this.typesenseIndex.count();
+    return 0;
+  }
+  /**
+   * Remove embeddings for the given node IDs from the active index.
+   * For cosine the `vectors` SQLite table is cleaned up automatically via FK cascade
+   * when nodes are deleted. For sqlite-vec, orama, and pglite the engine itself is
+   * the sole store of record, so we delete from it explicitly here.
+   */
+  async deleteEmbeddings(nodeIds) {
+    for (const id of nodeIds) {
+      this.vecIndex?.delete(id);
+      if (this.oramaIndex?.isAvailable()) await this.oramaIndex.delete(id);
+      if (this.pgliteIndex?.isAvailable()) await this.pgliteIndex.delete(id);
+      if (this.lancedbIndex?.isAvailable()) await this.lancedbIndex.delete(id);
+      if (this.qdrantIndex?.isAvailable()) await this.qdrantIndex.delete(id);
+      if (this.typesenseIndex?.isAvailable()) await this.typesenseIndex.delete(id);
+    }
+  }
+  /**
+   * Embed all eligible nodes in the database that don't yet have embeddings.
+   * Streams nodes in pages to avoid loading the entire node set into memory —
+   * critical for large codebases (100K+ symbols) where a single getAllNodes()
+   * call can exhaust the Node.js heap or the WASM linear memory.
+   *
+   * Uses dynamic batch sizing: starts at BATCH_SIZE, halves on OOM errors,
+   * and recovers back to full size after successful batches.
+   *
+   * Emits a pre-flight warning via onProgress when the embeddable node count
+   * exceeds LARGE_CODEBASE_THRESHOLD so the CLI can surface it to the user.
+   */
+  async embedAll(onProgress) {
+    if (!this.isInitialized() || !this.pipeline) return 0;
+    const modelId = this.config.embeddingModel || DEFAULT_MODEL;
+    const EMBEDDABLE_KINDS_ARRAY = [...EMBEDDABLE_KINDS];
+    const totalEmbeddable = this.db.countEmbeddableNodes(EMBEDDABLE_KINDS_ARRAY);
+    const LARGE_CODEBASE_THRESHOLD = 1e5;
+    if (totalEmbeddable > LARGE_CODEBASE_THRESHOLD) {
+      onProgress?.(-1, totalEmbeddable);
+    }
+    const existingIds = new Set(
+      this.typesenseIndex?.isAvailable() ? await this.typesenseIndex.getEmbeddedNodeIds() : this.qdrantIndex?.isAvailable() ? await this.qdrantIndex.getEmbeddedNodeIds() : this.lancedbIndex?.isAvailable() ? await this.lancedbIndex.getEmbeddedNodeIds() : this.pgliteIndex?.isAvailable() ? await this.pgliteIndex.getEmbeddedNodeIds() : this.oramaIndex?.isAvailable() ? await this.oramaIndex.getEmbeddedNodeIds() : this.vecIndex?.isAvailable() ? this.vecIndex.getEmbeddedNodeIds() : this.db.getEmbeddedNodeIds()
+    );
+    const PAGE_SIZE = 2e3;
+    let pageOffset = 0;
+    let processed = 0;
+    let currentBatchSize = BATCH_SIZE;
+    let consecutiveSuccesses = 0;
+    const totalPending = Math.max(0, totalEmbeddable - existingIds.size);
+    if (totalPending === 0) {
+      if (this.oramaIndex?.isAvailable()) await this.oramaIndex.save();
+      return 0;
+    }
+    (0, import_errors.logDebug)(`VectorManager: embedding ~${totalPending} nodes (${existingIds.size} already embedded, streaming in pages of ${PAGE_SIZE})`);
+    let pageBuffer = [];
+    let pageBufferOffset = 0;
+    let pageExhausted = false;
+    const fetchNextPage = () => {
+      const page = this.db.getEmbeddableNodesPaged(EMBEDDABLE_KINDS_ARRAY, PAGE_SIZE, pageOffset);
+      pageOffset += page.length;
+      pageBuffer = page.filter((n) => !existingIds.has(n.id));
+      pageBufferOffset = 0;
+      if (page.length < PAGE_SIZE) pageExhausted = true;
+    };
+    fetchNextPage();
+    while (pageBuffer.length > 0 || !pageExhausted && pageBufferOffset >= pageBuffer.length) {
+      if (pageBufferOffset >= pageBuffer.length) {
+        if (pageExhausted) break;
+        fetchNextPage();
+        if (pageBuffer.length === 0) break;
+      }
+      const batch = pageBuffer.slice(pageBufferOffset, pageBufferOffset + currentBatchSize);
+      if (batch.length === 0) break;
+      const texts = batch.map((n) => `search_document: ${nodeToText(n)}`);
+      try {
+        const outputs = await this.pipeline(texts, { pooling: "mean", normalize: true });
+        const dims = outputs.dims;
+        const dim = dims[1] ?? (this.config.embeddingDim ?? DEFAULT_EMBEDDING_DIM);
+        const flat = toFloat32Array(outputs.data);
+        const tsNodes = [];
+        const tsEmbeddings = [];
+        const qdNodes = [];
+        const qdEmbeddings = [];
+        for (let j = 0; j < batch.length; j++) {
+          const node = batch[j];
+          const embedding = flat.slice(j * dim, (j + 1) * dim);
+          if (!this.vecIndex?.isAvailable() && !this.oramaIndex?.isAvailable() && !this.pgliteIndex?.isAvailable() && !this.lancedbIndex?.isAvailable() && !this.qdrantIndex?.isAvailable() && !this.typesenseIndex?.isAvailable()) {
+            this.db.storeEmbedding(node.id, embedding, modelId);
+          }
+          this.vecIndex?.upsert(node.id, embedding);
+          if (this.oramaIndex?.isAvailable()) await this.oramaIndex.upsert(node, embedding);
+          if (this.pgliteIndex?.isAvailable()) await this.pgliteIndex.upsert(node, embedding);
+          if (this.lancedbIndex?.isAvailable()) await this.lancedbIndex.upsert(node, embedding);
+          if (this.qdrantIndex?.isAvailable()) {
+            qdNodes.push(node);
+            qdEmbeddings.push(embedding);
+          }
+          if (this.typesenseIndex?.isAvailable()) {
+            tsNodes.push(node);
+            tsEmbeddings.push(embedding);
+          }
+        }
+        if (qdNodes.length > 0) await this.qdrantIndex.bulkUpsert(qdNodes, qdEmbeddings);
+        if (tsNodes.length > 0) await this.typesenseIndex.bulkUpsert(tsNodes, tsEmbeddings);
+        pageBufferOffset += batch.length;
+        processed += batch.length;
+        consecutiveSuccesses++;
+        if (currentBatchSize < BATCH_SIZE && consecutiveSuccesses >= 5) {
+          currentBatchSize = Math.min(currentBatchSize * 2, BATCH_SIZE);
+          (0, import_errors.logDebug)(`VectorManager: batch size recovered to ${currentBatchSize}`);
+          consecutiveSuccesses = 0;
+        }
+      } catch (err) {
+        const errMsg = String(err);
+        const isOOM = errMsg.includes("bad allocation") || errMsg.includes("OOM") || errMsg.includes("out of memory");
+        if (isOOM && currentBatchSize > MIN_BATCH_SIZE) {
+          currentBatchSize = Math.max(Math.floor(currentBatchSize / 2), MIN_BATCH_SIZE);
+          consecutiveSuccesses = 0;
+          (0, import_errors.logWarn)(`VectorManager: OOM at batch offset ${pageBufferOffset} \u2014 reducing batch size to ${currentBatchSize} and retrying`, {
+            batchSize: currentBatchSize,
+            nodeIds: batch.slice(0, 3).map((n) => n.id)
+          });
+        } else if (isOOM && currentBatchSize <= MIN_BATCH_SIZE) {
+          (0, import_errors.logError)(`VectorManager: OOM even at minimum batch size ${MIN_BATCH_SIZE} \u2014 skipping ${batch.length} nodes`, {
+            batchStart: pageBufferOffset,
+            skippedNodes: batch.map((n) => `${n.kind}:${n.name} (${n.filePath})`).slice(0, 5)
+          });
+          pageBufferOffset += batch.length;
+          processed += batch.length;
+          consecutiveSuccesses = 0;
+        } else {
+          (0, import_errors.logWarn)(`VectorManager: batch embedding failed at offset ${pageBufferOffset}`, {
+            batchSize: currentBatchSize,
+            error: errMsg
+          });
+          pageBufferOffset += batch.length;
+          processed += batch.length;
+          consecutiveSuccesses = 0;
+        }
+      }
+      onProgress?.(processed, totalPending);
+    }
+    if (this.oramaIndex?.isAvailable()) await this.oramaIndex.save();
+    (0, import_errors.logDebug)(`VectorManager: embedding complete \u2014 ${processed} processed, final batch size: ${currentBatchSize}`);
+    return processed;
+  }
+  /**
+   * Semantic search: embed the query and return top-N nodes by similarity.
+   *
+   * When sqlite-vec is available (useVecIndex: true), delegates ANN search to
+   * VecIndex (fast, sub-linear). Otherwise falls back to in-process cosine
+   * similarity over all stored embeddings (linear scan, no extra deps).
+   *
+   * Returns empty array when not initialized.
+   */
+  async search(query, topN = 10) {
+    if (!this.isInitialized() || !this.pipeline) return [];
+    try {
+      const text = `search_query: ${query}`;
+      const output = await this.pipeline(text, { pooling: "mean", normalize: true });
+      const queryVec = toFloat32Array(output.data);
+      let nodeIds;
+      if (this.pgliteIndex?.isAvailable()) {
+        nodeIds = await this.pgliteIndex.search(query, queryVec, topN);
+      } else if (this.oramaIndex?.isAvailable()) {
+        nodeIds = await this.oramaIndex.search(query, queryVec, topN);
+      } else if (this.lancedbIndex?.isAvailable()) {
+        nodeIds = await this.lancedbIndex.search(queryVec, topN);
+      } else if (this.qdrantIndex?.isAvailable()) {
+        nodeIds = await this.qdrantIndex.search(queryVec, topN);
+      } else if (this.typesenseIndex?.isAvailable()) {
+        nodeIds = await this.typesenseIndex.search(queryVec, topN);
+      } else if (this.vecIndex?.isAvailable()) {
+        nodeIds = this.vecIndex.search(queryVec, topN);
+      } else {
+        const allEmbeddings = this.db.getAllEmbeddings();
+        nodeIds = allEmbeddings.map(({ nodeId, embedding }) => ({ nodeId, score: cosine(queryVec, embedding) })).filter((r) => r.score >= 0.3).sort((a, b) => b.score - a.score).slice(0, topN).map((r) => r.nodeId);
+      }
+      const results = [];
+      for (const nodeId of nodeIds) {
+        const node = this.db.getNode(nodeId);
+        if (node) results.push(node);
+      }
+      return results;
+    } catch (err) {
+      (0, import_errors.logWarn)("VectorManager: search failed", { error: String(err) });
+      return [];
+    }
+  }
+  /** Release engine resources (e.g. kill Qdrant child process). */
+  close() {
+    this.qdrantIndex?.close();
+    this.typesenseIndex?.close();
+  }
+}
+// Annotate the CommonJS export names for ESM import in node:
+0 && (module.exports = {
+  VectorManager
+});
+//# sourceMappingURL=index.js.map

package/dist/vectors/index.js.map

@@ -0,0 +1,7 @@
+{
+  "version": 3,
+  "sources": ["../../src/vectors/index.ts"],
+  "sourcesContent": ["/**\n * KiroGraph Vector Manager\n *\n * Semantic search via configurable local embedding models using @huggingface/transformers v3.\n * Mirrors CodeGraph src/vectors/ (embedder.ts + manager.ts) adapted for KiroGraph:\n *   - Cache dir: ~/.kirograph/models/ (not ~/.codegraph/models/)\n *   - Embeddings stored in the `vectors` SQLite table\n *   - Cosine similarity search done in-process by default (no extra deps)\n *   - ANN search via sqlite-vec opt-in: set config.useVecIndex = true\n *     (requires `npm install better-sqlite3 sqlite-vec`)\n *   - Disabled by default; opt-in via config.enableEmbeddings = true\n */\n\nimport * as path from 'path';\nimport * as fs from 'fs';\nimport { homedir } from 'os';\nimport { logDebug, logWarn, logError } from '../errors';\nimport type { KiroGraphConfig } from '../config';\nimport type { Node } from '../types';\nimport type { GraphDatabase } from '../db/database';\nimport { VecIndex } from './vec-index';\nimport { OramaIndex } from './orama-index';\nimport { PGliteIndex } from './pglite-index';\nimport { LanceDBIndex } from './lancedb-index';\nimport { QdrantIndex } from './qdrant-index';\nimport { TypesenseIndex } from './typesense-index';\n\n// \u2500\u2500 Constants \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n\nconst DEFAULT_MODEL = 'nomic-ai/nomic-embed-text-v1.5';\nconst DEFAULT_EMBEDDING_DIM = 768;\nconst GLOBAL_MODELS_DIR = path.join(homedir(), '.kirograph', 'models');\nconst BATCH_SIZE = 32;\nconst MIN_BATCH_SIZE = 4;\nconst MAX_TOKEN_CHARS = 2000;\n\n/** Node kinds worth embedding \u2014 high information density */\nconst EMBEDDABLE_KINDS = new Set<Node['kind']>([\n  'function', 'method', 'class', 'interface', 'type_alias', 'component', 'module',\n]);\n\n// \u2500\u2500 Embedder \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n\ntype Pipeline = any;\nlet transformers: typeof import('@huggingface/transformers') | null = null;\n\nasync function getTransformers() {\n  if (!transformers) {\n    transformers = await import('@huggingface/transformers');\n  }\n  return transformers;\n}\n\n/**\n * Build a searchable text representation of a node.\n * Mirrors CodeGraph TextEmbedder.createNodeText().\n * Truncated to MAX_TOKEN_CHARS to stay within the model's memory-safe token limit.\n * Priority: kind + name > qualified name > file path > signature > docstring.\n * This ensures the most important search signal is always preserved.\n */\nfunction nodeToText(node: Node): string {\n  const parts = [`${node.kind}: ${node.name}`];\n  if (node.qualifiedName && node.qualifiedName !== node.name) {\n    parts.push(`path: ${node.qualifiedName}`);\n  }\n  parts.push(`file: ${node.filePath}`);\n  if (node.signature) parts.push(`signature: ${node.signature}`);\n  if (node.docstring) parts.push(`documentation: ${node.docstring}`);\n  const text = parts.join('\\n');\n  if (text.length > MAX_TOKEN_CHARS) {\n    logDebug(`nodeToText: truncated ${node.id} from ${text.length} to ${MAX_TOKEN_CHARS} chars`);\n    return text.slice(0, MAX_TOKEN_CHARS);\n  }\n  return text;\n}\n\n/** Cosine similarity between two equal-length Float32Arrays. */\nfunction cosine(a: Float32Array, b: Float32Array): number {\n  let dot = 0, na = 0, nb = 0;\n  for (let i = 0; i < a.length; i++) {\n    dot += a[i]! * b[i]!;\n    na += a[i]! * a[i]!;\n    nb += b[i]! * b[i]!;\n  }\n  const denom = Math.sqrt(na) * Math.sqrt(nb);\n  return denom === 0 ? 0 : dot / denom;\n}\n\n/** Extract a flat Float32Array from the transformer pipeline output. */\nfunction toFloat32Array(data: unknown): Float32Array {\n  if (data instanceof Float32Array) return data;\n  if (Array.isArray(data)) return new Float32Array(data);\n  if (data && typeof data === 'object' && 'length' in data) {\n    return Float32Array.from(Array.from(data as ArrayLike<number>));\n  }\n  throw new Error('Unsupported embedding data format');\n}\n\n// \u2500\u2500 VectorManager \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n\nexport class VectorManager {\n  private pipeline: Pipeline | null = null;\n  private _initialized = false;\n  private vecIndex: VecIndex | null = null;\n  private oramaIndex: OramaIndex | null = null;\n  private pgliteIndex: PGliteIndex | null = null;\n  private lancedbIndex: LanceDBIndex | null = null;\n  private qdrantIndex: QdrantIndex | null = null;\n  private typesenseIndex: TypesenseIndex | null = null;\n  private _engineFallback: string | null = null;\n\n  constructor(\n    private readonly db: GraphDatabase,\n    private readonly config: KiroGraphConfig,\n    private readonly projectRoot?: string,\n  ) {}\n\n  isInitialized(): boolean {\n    return this.config.enableEmbeddings === true && this._initialized;\n  }\n\n  /** Non-null when the configured engine failed to load and cosine is being used instead. */\n  getEngineFallback(): string | null {\n    return this._engineFallback;\n  }\n\n  /**\n   * Load the embedding model. No-op when embeddings are disabled.\n   * Fails silently so callers can continue without semantic search.\n   * When config.useVecIndex is true, also initializes the sqlite-vec ANN index.\n   */\n  async initialize(onProgress?: (file: string, loaded: number, total: number, done: boolean) => void): Promise<void> {\n    if (!this.config.enableEmbeddings) {\n      logDebug('VectorManager: embeddings disabled');\n      return;\n    }\n    if (this._initialized) return;\n\n    const rawModel = this.config.embeddingModel || DEFAULT_MODEL;\n    // Guard against an invalid model ID (e.g. 'y' typed by mistake in the installer)\n    const modelId = rawModel.includes('/') ? rawModel : DEFAULT_MODEL;\n    if (rawModel !== modelId) {\n      logWarn(`VectorManager: invalid embeddingModel \"${rawModel}\", using default instead`, { default: DEFAULT_MODEL });\n    }\n    const cacheDir = GLOBAL_MODELS_DIR;\n\n    try {\n      const { pipeline, env } = await getTransformers();\n      env.cacheDir = cacheDir;\n\n      if (!fs.existsSync(cacheDir)) {\n        fs.mkdirSync(cacheDir, { recursive: true });\n      }\n\n      // Skip remote check if model already cached (HuggingFace uses '--' as path separator)\n      // HuggingFace transformers v3 stores models as <org>/<model>/ (v2 used <org>--<model>)\n      const cached = fs.existsSync(path.join(cacheDir, modelId));\n      if (cached) env.allowRemoteModels = false;\n\n      let lastFile = '';\n      this.pipeline = await pipeline('feature-extraction', modelId, {\n        dtype: 'auto',\n        ...(!cached && onProgress ? {\n          progress_callback: (p: { status: string; file?: string; loaded?: number; total?: number }) => {\n            if (p.status === 'progress' && p.file) {\n              lastFile = p.file;\n              onProgress(p.file, p.loaded ?? 0, p.total ?? 0, false);\n            } else if (p.status === 'done' && lastFile) {\n              onProgress(lastFile, 1, 1, true);\n              lastFile = '';\n            }\n          },\n        } : {}),\n      });\n\n      // Validate embeddingDim against the model's actual output shape.\n      // Run a single test embedding and compare dims[1] to config.embeddingDim.\n      // If they differ, warn and update the runtime dim so downstream engines\n      // are initialised with the correct size (config.json is not rewritten here).\n      try {\n        const testOut = await this.pipeline('test', { pooling: 'mean', normalize: true });\n        const actualDim: number = testOut.dims?.[1] ?? (testOut.data?.length as number);\n        const configuredDim = this.config.embeddingDim ?? DEFAULT_EMBEDDING_DIM;\n        if (actualDim && actualDim !== configuredDim) {\n          logWarn(\n            `VectorManager: embeddingDim mismatch \u2014 config says ${configuredDim} but model outputs ${actualDim}. Using actual dim.`,\n            { model: modelId, configuredDim, actualDim },\n          );\n          // Patch the runtime config so all engine constructors below use the correct dim\n          (this.config as unknown as Record<string, unknown>).embeddingDim = actualDim;\n        }\n      } catch { /* test embedding failed \u2014 proceed with configured dim */ }\n\n      this._initialized = true;\n      logDebug('VectorManager: model loaded', { modelId, embeddingDim: this.config.embeddingDim ?? DEFAULT_EMBEDDING_DIM });\n    } catch (err) {\n      logError('VectorManager: failed to load embedding model', {\n        model: modelId,\n        error: String(err),\n      });\n      this._initialized = false;\n      return;\n    }\n\n    // Initialize the selected search engine\n    const engine = this.config.semanticEngine ?? (this.config.useVecIndex ? 'sqlite-vec' : 'cosine');\n\n    if (this.projectRoot && engine !== 'cosine') {\n      const kirographDir = path.join(this.projectRoot, '.kirograph');\n\n      const embeddingDim = this.config.embeddingDim ?? DEFAULT_EMBEDDING_DIM;\n\n    if (engine === 'sqlite-vec') {\n        this.vecIndex = new VecIndex(kirographDir, embeddingDim);\n        await this.vecIndex.initialize();\n        if (this.vecIndex.isAvailable()) {\n          logDebug('VectorManager: sqlite-vec ANN index ready');\n        } else {\n          this._engineFallback = 'sqlite-vec unavailable \u2014 run: npm install better-sqlite3 sqlite-vec';\n          logDebug('VectorManager: sqlite-vec unavailable, falling back to in-process cosine');\n        }\n      } else if (engine === 'orama') {\n        this.oramaIndex = new OramaIndex(kirographDir, embeddingDim);\n        await this.oramaIndex.initialize();\n        if (this.oramaIndex.isAvailable()) {\n          logDebug('VectorManager: Orama hybrid index ready');\n        } else {\n          this._engineFallback = 'orama unavailable \u2014 run: npm install @orama/orama @orama/plugin-data-persistence';\n          logDebug('VectorManager: Orama unavailable, falling back to in-process cosine');\n        }\n      } else if (engine === 'pglite') {\n        this.pgliteIndex = new PGliteIndex(kirographDir, embeddingDim);\n        await this.pgliteIndex.initialize();\n        if (this.pgliteIndex.isAvailable()) {\n          logDebug('VectorManager: PGlite+pgvector hybrid index ready');\n        } else {\n          this._engineFallback = 'pglite unavailable \u2014 run: npm install @electric-sql/pglite';\n          logDebug('VectorManager: PGlite unavailable, falling back to in-process cosine');\n        }\n      } else if (engine === 'lancedb') {\n        this.lancedbIndex = new LanceDBIndex(kirographDir, embeddingDim);\n        await this.lancedbIndex.initialize();\n        if (this.lancedbIndex.isAvailable()) {\n          logDebug('VectorManager: LanceDB ANN index ready');\n        } else {\n          this._engineFallback = 'lancedb unavailable \u2014 run: npm install @lancedb/lancedb';\n          logDebug('VectorManager: LanceDB unavailable, falling back to in-process cosine');\n        }\n      } else if (engine === 'qdrant') {\n        this.qdrantIndex = new QdrantIndex(kirographDir, embeddingDim);\n        await this.qdrantIndex.initialize();\n        if (this.qdrantIndex.isAvailable()) {\n          logDebug('VectorManager: Qdrant ANN index ready');\n        } else {\n          this._engineFallback = 'qdrant unavailable \u2014 run: npm install qdrant-local';\n          logDebug('VectorManager: Qdrant unavailable, falling back to in-process cosine');\n        }\n      } else if (engine === 'typesense') {\n        this.typesenseIndex = new TypesenseIndex(kirographDir, embeddingDim);\n        await this.typesenseIndex.initialize();\n        if (this.typesenseIndex.isAvailable()) {\n          logDebug('VectorManager: Typesense ANN index ready');\n        } else {\n          const reason = this.typesenseIndex.getFailReason();\n          this._engineFallback = reason\n            ? `typesense initialization failed: ${reason}`\n            : 'typesense unavailable \u2014 run: npm install typesense';\n          logDebug('VectorManager: Typesense unavailable, falling back to in-process cosine');\n        }\n      }\n    }\n  }\n\n  /**\n   * Embed a single node and persist to the vectors table.\n   * Skips silently when disabled; logs on failure without throwing.\n   */\n  async embedNode(node: Node): Promise<void> {\n    if (!this.config.enableEmbeddings) return;\n    if (!this._initialized || !this.pipeline) {\n      logError('Embedding model unavailable', { model: this.config.embeddingModel });\n      return;\n    }\n    if (!EMBEDDABLE_KINDS.has(node.kind)) return;\n\n    try {\n      const text = `search_document: ${nodeToText(node)}`;\n      const output = await this.pipeline(text, { pooling: 'mean', normalize: true });\n      const embedding = toFloat32Array(output.data);\n      // non-cosine engines are sole stores of record when active \u2014 skip the SQLite vectors table\n      if (!this.vecIndex?.isAvailable() && !this.oramaIndex?.isAvailable() && !this.pgliteIndex?.isAvailable() && !this.lancedbIndex?.isAvailable() && !this.qdrantIndex?.isAvailable() && !this.typesenseIndex?.isAvailable()) {\n        this.db.storeEmbedding(node.id, embedding, this.config.embeddingModel || DEFAULT_MODEL);\n      }\n      this.vecIndex?.upsert(node.id, embedding);\n      if (this.oramaIndex?.isAvailable()) await this.oramaIndex.upsert(node, embedding);\n      if (this.pgliteIndex?.isAvailable()) await this.pgliteIndex.upsert(node, embedding);\n      if (this.lancedbIndex?.isAvailable()) await this.lancedbIndex.upsert(node, embedding);\n      if (this.qdrantIndex?.isAvailable()) await this.qdrantIndex.upsert(node, embedding);\n      if (this.typesenseIndex?.isAvailable()) await this.typesenseIndex.upsert(node, embedding);\n    } catch (err) {\n      logWarn('Failed to embed node', { nodeId: node.id, error: String(err) });\n    }\n  }\n\n  /** Number of entries currently in the active ANN/hybrid index (0 when not in use). */\n  async vecIndexCount(): Promise<number> {\n    if (this.vecIndex?.isAvailable()) return this.vecIndex.count();\n    if (this.oramaIndex?.isAvailable()) return this.oramaIndex.count();\n    if (this.pgliteIndex?.isAvailable()) return this.pgliteIndex.count();\n    if (this.lancedbIndex?.isAvailable()) return this.lancedbIndex.count();\n    if (this.qdrantIndex?.isAvailable()) return this.qdrantIndex.count();\n    if (this.typesenseIndex?.isAvailable()) return this.typesenseIndex.count();\n    return 0;\n  }\n\n  /**\n   * Remove embeddings for the given node IDs from the active index.\n   * For cosine the `vectors` SQLite table is cleaned up automatically via FK cascade\n   * when nodes are deleted. For sqlite-vec, orama, and pglite the engine itself is\n   * the sole store of record, so we delete from it explicitly here.\n   */\n  async deleteEmbeddings(nodeIds: string[]): Promise<void> {\n    for (const id of nodeIds) {\n      this.vecIndex?.delete(id);\n      if (this.oramaIndex?.isAvailable()) await this.oramaIndex.delete(id);\n      if (this.pgliteIndex?.isAvailable()) await this.pgliteIndex.delete(id);\n      if (this.lancedbIndex?.isAvailable()) await this.lancedbIndex.delete(id);\n      if (this.qdrantIndex?.isAvailable()) await this.qdrantIndex.delete(id);\n      if (this.typesenseIndex?.isAvailable()) await this.typesenseIndex.delete(id);\n    }\n  }\n\n  /**\n   * Embed all eligible nodes in the database that don't yet have embeddings.\n   * Streams nodes in pages to avoid loading the entire node set into memory \u2014\n   * critical for large codebases (100K+ symbols) where a single getAllNodes()\n   * call can exhaust the Node.js heap or the WASM linear memory.\n   *\n   * Uses dynamic batch sizing: starts at BATCH_SIZE, halves on OOM errors,\n   * and recovers back to full size after successful batches.\n   *\n   * Emits a pre-flight warning via onProgress when the embeddable node count\n   * exceeds LARGE_CODEBASE_THRESHOLD so the CLI can surface it to the user.\n   */\n  async embedAll(onProgress?: (current: number, total: number) => void): Promise<number> {\n    if (!this.isInitialized() || !this.pipeline) return 0;\n\n    const modelId = this.config.embeddingModel || DEFAULT_MODEL;\n    const EMBEDDABLE_KINDS_ARRAY = [...EMBEDDABLE_KINDS] as string[];\n\n    // Pre-flight: count embeddable nodes without loading them\n    const totalEmbeddable = this.db.countEmbeddableNodes(EMBEDDABLE_KINDS_ARRAY);\n    const LARGE_CODEBASE_THRESHOLD = 100_000;\n    if (totalEmbeddable > LARGE_CODEBASE_THRESHOLD) {\n      // Signal the large-codebase warning via a special progress event.\n      // current=-1 is the sentinel; the CLI renderer checks for it.\n      onProgress?.(-1, totalEmbeddable);\n    }\n\n    // Collect already-embedded IDs (needed to skip nodes already in the index)\n    const existingIds = new Set(\n      this.typesenseIndex?.isAvailable()\n        ? await this.typesenseIndex.getEmbeddedNodeIds()\n        : this.qdrantIndex?.isAvailable()\n        ? await this.qdrantIndex.getEmbeddedNodeIds()\n        : this.lancedbIndex?.isAvailable()\n        ? await this.lancedbIndex.getEmbeddedNodeIds()\n        : this.pgliteIndex?.isAvailable()\n          ? await this.pgliteIndex.getEmbeddedNodeIds()\n          : this.oramaIndex?.isAvailable()\n            ? await this.oramaIndex.getEmbeddedNodeIds()\n            : this.vecIndex?.isAvailable()\n              ? this.vecIndex.getEmbeddedNodeIds()\n              : this.db.getEmbeddedNodeIds()\n    );\n\n    // Stream nodes in pages \u2014 never hold more than PAGE_SIZE nodes in memory at once\n    const PAGE_SIZE = 2000;\n    let pageOffset = 0;\n    let processed = 0;\n    let currentBatchSize = BATCH_SIZE;\n    let consecutiveSuccesses = 0;\n\n    // We need to know the total pending count for progress reporting.\n    // Approximate: totalEmbeddable - existingIds.size (may be slightly off if\n    // existingIds contains IDs for non-embeddable kinds, but close enough).\n    const totalPending = Math.max(0, totalEmbeddable - existingIds.size);\n    if (totalPending === 0) {\n      if (this.oramaIndex?.isAvailable()) await this.oramaIndex.save();\n      return 0;\n    }\n\n    logDebug(`VectorManager: embedding ~${totalPending} nodes (${existingIds.size} already embedded, streaming in pages of ${PAGE_SIZE})`);\n\n    // Buffer of pending nodes from the current page\n    let pageBuffer: Node[] = [];\n    let pageBufferOffset = 0; // index within pageBuffer\n    let pageExhausted = false;\n\n    const fetchNextPage = () => {\n      const page = this.db.getEmbeddableNodesPaged(EMBEDDABLE_KINDS_ARRAY, PAGE_SIZE, pageOffset);\n      pageOffset += page.length;\n      pageBuffer = page.filter(n => !existingIds.has(n.id));\n      pageBufferOffset = 0;\n      if (page.length < PAGE_SIZE) pageExhausted = true;\n    };\n\n    fetchNextPage();\n\n    while (pageBuffer.length > 0 || (!pageExhausted && pageBufferOffset >= pageBuffer.length)) {\n      // Refill page buffer when exhausted\n      if (pageBufferOffset >= pageBuffer.length) {\n        if (pageExhausted) break;\n        fetchNextPage();\n        if (pageBuffer.length === 0) break;\n      }\n\n      // Slice a batch from the current page buffer\n      const batch = pageBuffer.slice(pageBufferOffset, pageBufferOffset + currentBatchSize);\n      if (batch.length === 0) break;\n\n      const texts = batch.map(n => `search_document: ${nodeToText(n)}`);\n\n      try {\n        const outputs = await this.pipeline(texts, { pooling: 'mean', normalize: true });\n        const dims: number[] = outputs.dims;\n        const dim = dims[1] ?? (this.config.embeddingDim ?? DEFAULT_EMBEDDING_DIM);\n        const flat = toFloat32Array(outputs.data);\n\n        const tsNodes: Node[] = [];\n        const tsEmbeddings: Float32Array[] = [];\n        const qdNodes: Node[] = [];\n        const qdEmbeddings: Float32Array[] = [];\n\n        for (let j = 0; j < batch.length; j++) {\n          const node = batch[j]!;\n          const embedding = flat.slice(j * dim, (j + 1) * dim);\n          if (!this.vecIndex?.isAvailable() && !this.oramaIndex?.isAvailable() && !this.pgliteIndex?.isAvailable() && !this.lancedbIndex?.isAvailable() && !this.qdrantIndex?.isAvailable() && !this.typesenseIndex?.isAvailable()) {\n            this.db.storeEmbedding(node.id, embedding, modelId);\n          }\n          this.vecIndex?.upsert(node.id, embedding);\n          if (this.oramaIndex?.isAvailable()) await this.oramaIndex.upsert(node, embedding);\n          if (this.pgliteIndex?.isAvailable()) await this.pgliteIndex.upsert(node, embedding);\n          if (this.lancedbIndex?.isAvailable()) await this.lancedbIndex.upsert(node, embedding);\n          if (this.qdrantIndex?.isAvailable()) { qdNodes.push(node); qdEmbeddings.push(embedding); }\n          if (this.typesenseIndex?.isAvailable()) { tsNodes.push(node); tsEmbeddings.push(embedding); }\n        }\n\n        if (qdNodes.length > 0) await this.qdrantIndex!.bulkUpsert(qdNodes, qdEmbeddings);\n        if (tsNodes.length > 0) await this.typesenseIndex!.bulkUpsert(tsNodes, tsEmbeddings);\n\n        pageBufferOffset += batch.length;\n        processed += batch.length;\n        consecutiveSuccesses++;\n\n        if (currentBatchSize < BATCH_SIZE && consecutiveSuccesses >= 5) {\n          currentBatchSize = Math.min(currentBatchSize * 2, BATCH_SIZE);\n          logDebug(`VectorManager: batch size recovered to ${currentBatchSize}`);\n          consecutiveSuccesses = 0;\n        }\n      } catch (err) {\n        const errMsg = String(err);\n        const isOOM = errMsg.includes('bad allocation') || errMsg.includes('OOM') || errMsg.includes('out of memory');\n\n        if (isOOM && currentBatchSize > MIN_BATCH_SIZE) {\n          currentBatchSize = Math.max(Math.floor(currentBatchSize / 2), MIN_BATCH_SIZE);\n          consecutiveSuccesses = 0;\n          logWarn(`VectorManager: OOM at batch offset ${pageBufferOffset} \u2014 reducing batch size to ${currentBatchSize} and retrying`, {\n            batchSize: currentBatchSize,\n            nodeIds: batch.slice(0, 3).map(n => n.id),\n          });\n          // Don't advance pageBufferOffset \u2014 retry the same batch\n        } else if (isOOM && currentBatchSize <= MIN_BATCH_SIZE) {\n          logError(`VectorManager: OOM even at minimum batch size ${MIN_BATCH_SIZE} \u2014 skipping ${batch.length} nodes`, {\n            batchStart: pageBufferOffset,\n            skippedNodes: batch.map(n => `${n.kind}:${n.name} (${n.filePath})`).slice(0, 5),\n          });\n          pageBufferOffset += batch.length;\n          processed += batch.length;\n          consecutiveSuccesses = 0;\n        } else {\n          logWarn(`VectorManager: batch embedding failed at offset ${pageBufferOffset}`, {\n            batchSize: currentBatchSize,\n            error: errMsg,\n          });\n          pageBufferOffset += batch.length;\n          processed += batch.length;\n          consecutiveSuccesses = 0;\n        }\n      }\n\n      onProgress?.(processed, totalPending);\n    }\n\n    if (this.oramaIndex?.isAvailable()) await this.oramaIndex.save();\n\n    logDebug(`VectorManager: embedding complete \u2014 ${processed} processed, final batch size: ${currentBatchSize}`);\n\n    return processed;\n  }\n\n  /**\n   * Semantic search: embed the query and return top-N nodes by similarity.\n   *\n   * When sqlite-vec is available (useVecIndex: true), delegates ANN search to\n   * VecIndex (fast, sub-linear). Otherwise falls back to in-process cosine\n   * similarity over all stored embeddings (linear scan, no extra deps).\n   *\n   * Returns empty array when not initialized.\n   */\n  async search(query: string, topN = 10): Promise<Node[]> {\n    if (!this.isInitialized() || !this.pipeline) return [];\n\n    try {\n      const text = `search_query: ${query}`;\n      const output = await this.pipeline(text, { pooling: 'mean', normalize: true });\n      const queryVec = toFloat32Array(output.data);\n\n      let nodeIds: string[];\n\n      if (this.pgliteIndex?.isAvailable()) {\n        // Hybrid search via PGlite+pgvector \u2014 exact vector + full-text in one SQL query\n        nodeIds = await this.pgliteIndex.search(query, queryVec, topN);\n      } else if (this.oramaIndex?.isAvailable()) {\n        // Hybrid search via Orama \u2014 full-text + vector combined\n        nodeIds = await this.oramaIndex.search(query, queryVec, topN);\n      } else if (this.lancedbIndex?.isAvailable()) {\n        // ANN search via LanceDB \u2014 columnar Lance format, cosine metric\n        nodeIds = await this.lancedbIndex.search(queryVec, topN);\n      } else if (this.qdrantIndex?.isAvailable()) {\n        // ANN search via Qdrant \u2014 HNSW index, cosine metric\n        nodeIds = await this.qdrantIndex.search(queryVec, topN);\n      } else if (this.typesenseIndex?.isAvailable()) {\n        // ANN search via Typesense \u2014 HNSW index, cosine metric\n        nodeIds = await this.typesenseIndex.search(queryVec, topN);\n      } else if (this.vecIndex?.isAvailable()) {\n        // ANN search via sqlite-vec \u2014 fast, sub-linear\n        nodeIds = this.vecIndex.search(queryVec, topN);\n      } else {\n        // In-process cosine similarity \u2014 linear scan over all stored embeddings\n        const allEmbeddings = this.db.getAllEmbeddings();\n        nodeIds = allEmbeddings\n          .map(({ nodeId, embedding }) => ({ nodeId, score: cosine(queryVec, embedding) }))\n          .filter(r => r.score >= 0.3)\n          .sort((a, b) => b.score - a.score)\n          .slice(0, topN)\n          .map(r => r.nodeId);\n      }\n\n      const results: Node[] = [];\n      for (const nodeId of nodeIds) {\n        const node = this.db.getNode(nodeId);\n        if (node) results.push(node);\n      }\n      return results;\n    } catch (err) {\n      logWarn('VectorManager: search failed', { error: String(err) });\n      return [];\n    }\n  }\n\n  /** Release engine resources (e.g. kill Qdrant child process). */\n  close(): void {\n    this.qdrantIndex?.close();\n    this.typesenseIndex?.close();\n  }\n}\n"],
+  "mappings": ";;;;;;;;;;;;;;;;;;;;;;;;;;;;AAAA;AAAA;AAAA;AAAA;AAAA;AAaA,WAAsB;AACtB,SAAoB;AACpB,gBAAwB;AACxB,oBAA4C;AAI5C,uBAAyB;AACzB,yBAA2B;AAC3B,0BAA4B;AAC5B,2BAA6B;AAC7B,0BAA4B;AAC5B,6BAA+B;AAI/B,MAAM,gBAAgB;AACtB,MAAM,wBAAwB;AAC9B,MAAM,oBAAoB,KAAK,SAAK,mBAAQ,GAAG,cAAc,QAAQ;AACrE,MAAM,aAAa;AACnB,MAAM,iBAAiB;AACvB,MAAM,kBAAkB;AAGxB,MAAM,mBAAmB,oBAAI,IAAkB;AAAA,EAC7C;AAAA,EAAY;AAAA,EAAU;AAAA,EAAS;AAAA,EAAa;AAAA,EAAc;AAAA,EAAa;AACzE,CAAC;AAKD,IAAI,eAAkE;AAEtE,eAAe,kBAAkB;AAC/B,MAAI,CAAC,cAAc;AACjB,mBAAe,MAAM,OAAO,2BAA2B;AAAA,EACzD;AACA,SAAO;AACT;AASA,SAAS,WAAW,MAAoB;AACtC,QAAM,QAAQ,CAAC,GAAG,KAAK,IAAI,KAAK,KAAK,IAAI,EAAE;AAC3C,MAAI,KAAK,iBAAiB,KAAK,kBAAkB,KAAK,MAAM;AAC1D,UAAM,KAAK,SAAS,KAAK,aAAa,EAAE;AAAA,EAC1C;AACA,QAAM,KAAK,SAAS,KAAK,QAAQ,EAAE;AACnC,MAAI,KAAK,UAAW,OAAM,KAAK,cAAc,KAAK,SAAS,EAAE;AAC7D,MAAI,KAAK,UAAW,OAAM,KAAK,kBAAkB,KAAK,SAAS,EAAE;AACjE,QAAM,OAAO,MAAM,KAAK,IAAI;AAC5B,MAAI,KAAK,SAAS,iBAAiB;AACjC,gCAAS,yBAAyB,KAAK,EAAE,SAAS,KAAK,MAAM,OAAO,eAAe,QAAQ;AAC3F,WAAO,KAAK,MAAM,GAAG,eAAe;AAAA,EACtC;AACA,SAAO;AACT;AAGA,SAAS,OAAO,GAAiB,GAAyB;AACxD,MAAI,MAAM,GAAG,KAAK,GAAG,KAAK;AAC1B,WAAS,IAAI,GAAG,IAAI,EAAE,QAAQ,KAAK;AACjC,WAAO,EAAE,CAAC,IAAK,EAAE,CAAC;AAClB,UAAM,EAAE,CAAC,IAAK,EAAE,CAAC;AACjB,UAAM,EAAE,CAAC,IAAK,EAAE,CAAC;AAAA,EACnB;AACA,QAAM,QAAQ,KAAK,KAAK,EAAE,IAAI,KAAK,KAAK,EAAE;AAC1C,SAAO,UAAU,IAAI,IAAI,MAAM;AACjC;AAGA,SAAS,eAAe,MAA6B;AACnD,MAAI,gBAAgB,aAAc,QAAO;AACzC,MAAI,MAAM,QAAQ,IAAI,EAAG,QAAO,IAAI,aAAa,IAAI;AACrD,MAAI,QAAQ,OAAO,SAAS,YAAY,YAAY,MAAM;AACxD,WAAO,aAAa,KAAK,MAAM,KAAK,IAAyB,CAAC;AAAA,EAChE;AACA,QAAM,IAAI,MAAM,mCAAmC;AACrD;AAIO,MAAM,cAAc;AAAA,EAWzB,YACmB,IACA,QACA,aACjB;AAHiB;AACA;AACA;AAbnB,SAAQ,WAA4B;AACpC,SAAQ,eAAe;AACvB,SAAQ,WAA4B;AACpC,SAAQ,aAAgC;AACxC,SAAQ,cAAkC;AAC1C,SAAQ,eAAoC;AAC5C,SAAQ,cAAkC;AAC1C,SAAQ,iBAAwC;AAChD,SAAQ,kBAAiC;AAAA,EAMtC;AAAA,EAEH,gBAAyB;AACvB,WAAO,KAAK,OAAO,qBAAqB,QAAQ,KAAK;AAAA,EACvD;AAAA;AAAA,EAGA,oBAAmC;AACjC,WAAO,KAAK;AAAA,EACd;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAOA,MAAM,WAAW,YAAkG;AACjH,QAAI,CAAC,KAAK,OAAO,kBAAkB;AACjC,kCAAS,oCAAoC;AAC7C;AAAA,IACF;AACA,QAAI,KAAK,aAAc;AAEvB,UAAM,WAAW,KAAK,OAAO,kBAAkB;AAE/C,UAAM,UAAU,SAAS,SAAS,GAAG,IAAI,WAAW;AACpD,QAAI,aAAa,SAAS;AACxB,iCAAQ,0CAA0C,QAAQ,4BAA4B,EAAE,SAAS,cAAc,CAAC;AAAA,IAClH;AACA,UAAM,WAAW;AAEjB,QAAI;AACF,YAAM,EAAE,UAAU,IAAI,IAAI,MAAM,gBAAgB;AAChD,UAAI,WAAW;AAEf,UAAI,CAAC,GAAG,WAAW,QAAQ,GAAG;AAC5B,WAAG,UAAU,UAAU,EAAE,WAAW,KAAK,CAAC;AAAA,MAC5C;AAIA,YAAM,SAAS,GAAG,WAAW,KAAK,KAAK,UAAU,OAAO,CAAC;AACzD,UAAI,OAAQ,KAAI,oBAAoB;AAEpC,UAAI,WAAW;AACf,WAAK,WAAW,MAAM,SAAS,sBAAsB,SAAS;AAAA,QAC5D,OAAO;AAAA,QACP,GAAI,CAAC,UAAU,aAAa;AAAA,UAC1B,mBAAmB,CAAC,MAA0E;AAC5F,gBAAI,EAAE,WAAW,cAAc,EAAE,MAAM;AACrC,yBAAW,EAAE;AACb,yBAAW,EAAE,MAAM,EAAE,UAAU,GAAG,EAAE,SAAS,GAAG,KAAK;AAAA,YACvD,WAAW,EAAE,WAAW,UAAU,UAAU;AAC1C,yBAAW,UAAU,GAAG,GAAG,IAAI;AAC/B,yBAAW;AAAA,YACb;AAAA,UACF;AAAA,QACF,IAAI,CAAC;AAAA,MACP,CAAC;AAMD,UAAI;AACF,cAAM,UAAU,MAAM,KAAK,SAAS,QAAQ,EAAE,SAAS,QAAQ,WAAW,KAAK,CAAC;AAChF,cAAM,YAAoB,QAAQ,OAAO,CAAC,KAAM,QAAQ,MAAM;AAC9D,cAAM,gBAAgB,KAAK,OAAO,gBAAgB;AAClD,YAAI,aAAa,cAAc,eAAe;AAC5C;AAAA,YACE,2DAAsD,aAAa,sBAAsB,SAAS;AAAA,YAClG,EAAE,OAAO,SAAS,eAAe,UAAU;AAAA,UAC7C;AAEA,UAAC,KAAK,OAA8C,eAAe;AAAA,QACrE;AAAA,MACF,QAAQ;AAAA,MAA4D;AAEpE,WAAK,eAAe;AACpB,kCAAS,+BAA+B,EAAE,SAAS,cAAc,KAAK,OAAO,gBAAgB,sBAAsB,CAAC;AAAA,IACtH,SAAS,KAAK;AACZ,kCAAS,iDAAiD;AAAA,QACxD,OAAO;AAAA,QACP,OAAO,OAAO,GAAG;AAAA,MACnB,CAAC;AACD,WAAK,eAAe;AACpB;AAAA,IACF;AAGA,UAAM,SAAS,KAAK,OAAO,mBAAmB,KAAK,OAAO,cAAc,eAAe;AAEvF,QAAI,KAAK,eAAe,WAAW,UAAU;AAC3C,YAAM,eAAe,KAAK,KAAK,KAAK,aAAa,YAAY;AAE7D,YAAM,eAAe,KAAK,OAAO,gBAAgB;AAEnD,UAAI,WAAW,cAAc;AACzB,aAAK,WAAW,IAAI,0BAAS,cAAc,YAAY;AACvD,cAAM,KAAK,SAAS,WAAW;AAC/B,YAAI,KAAK,SAAS,YAAY,GAAG;AAC/B,sCAAS,2CAA2C;AAAA,QACtD,OAAO;AACL,eAAK,kBAAkB;AACvB,sCAAS,0EAA0E;AAAA,QACrF;AAAA,MACF,WAAW,WAAW,SAAS;AAC7B,aAAK,aAAa,IAAI,8BAAW,cAAc,YAAY;AAC3D,cAAM,KAAK,WAAW,WAAW;AACjC,YAAI,KAAK,WAAW,YAAY,GAAG;AACjC,sCAAS,yCAAyC;AAAA,QACpD,OAAO;AACL,eAAK,kBAAkB;AACvB,sCAAS,qEAAqE;AAAA,QAChF;AAAA,MACF,WAAW,WAAW,UAAU;AAC9B,aAAK,cAAc,IAAI,gCAAY,cAAc,YAAY;AAC7D,cAAM,KAAK,YAAY,WAAW;AAClC,YAAI,KAAK,YAAY,YAAY,GAAG;AAClC,sCAAS,mDAAmD;AAAA,QAC9D,OAAO;AACL,eAAK,kBAAkB;AACvB,sCAAS,sEAAsE;AAAA,QACjF;AAAA,MACF,WAAW,WAAW,WAAW;AAC/B,aAAK,eAAe,IAAI,kCAAa,cAAc,YAAY;AAC/D,cAAM,KAAK,aAAa,WAAW;AACnC,YAAI,KAAK,aAAa,YAAY,GAAG;AACnC,sCAAS,wCAAwC;AAAA,QACnD,OAAO;AACL,eAAK,kBAAkB;AACvB,sCAAS,uEAAuE;AAAA,QAClF;AAAA,MACF,WAAW,WAAW,UAAU;AAC9B,aAAK,cAAc,IAAI,gCAAY,cAAc,YAAY;AAC7D,cAAM,KAAK,YAAY,WAAW;AAClC,YAAI,KAAK,YAAY,YAAY,GAAG;AAClC,sCAAS,uCAAuC;AAAA,QAClD,OAAO;AACL,eAAK,kBAAkB;AACvB,sCAAS,sEAAsE;AAAA,QACjF;AAAA,MACF,WAAW,WAAW,aAAa;AACjC,aAAK,iBAAiB,IAAI,sCAAe,cAAc,YAAY;AACnE,cAAM,KAAK,eAAe,WAAW;AACrC,YAAI,KAAK,eAAe,YAAY,GAAG;AACrC,sCAAS,0CAA0C;AAAA,QACrD,OAAO;AACL,gBAAM,SAAS,KAAK,eAAe,cAAc;AACjD,eAAK,kBAAkB,SACnB,oCAAoC,MAAM,KAC1C;AACJ,sCAAS,yEAAyE;AAAA,QACpF;AAAA,MACF;AAAA,IACF;AAAA,EACF;AAAA;AAAA;AAAA;AAAA;AAAA,EAMA,MAAM,UAAU,MAA2B;AACzC,QAAI,CAAC,KAAK,OAAO,iBAAkB;AACnC,QAAI,CAAC,KAAK,gBAAgB,CAAC,KAAK,UAAU;AACxC,kCAAS,+BAA+B,EAAE,OAAO,KAAK,OAAO,eAAe,CAAC;AAC7E;AAAA,IACF;AACA,QAAI,CAAC,iBAAiB,IAAI,KAAK,IAAI,EAAG;AAEtC,QAAI;AACF,YAAM,OAAO,oBAAoB,WAAW,IAAI,CAAC;AACjD,YAAM,SAAS,MAAM,KAAK,SAAS,MAAM,EAAE,SAAS,QAAQ,WAAW,KAAK,CAAC;AAC7E,YAAM,YAAY,eAAe,OAAO,IAAI;AAE5C,UAAI,CAAC,KAAK,UAAU,YAAY,KAAK,CAAC,KAAK,YAAY,YAAY,KAAK,CAAC,KAAK,aAAa,YAAY,KAAK,CAAC,KAAK,cAAc,YAAY,KAAK,CAAC,KAAK,aAAa,YAAY,KAAK,CAAC,KAAK,gBAAgB,YAAY,GAAG;AACxN,aAAK,GAAG,eAAe,KAAK,IAAI,WAAW,KAAK,OAAO,kBAAkB,aAAa;AAAA,MACxF;AACA,WAAK,UAAU,OAAO,KAAK,IAAI,SAAS;AACxC,UAAI,KAAK,YAAY,YAAY,EAAG,OAAM,KAAK,WAAW,OAAO,MAAM,SAAS;AAChF,UAAI,KAAK,aAAa,YAAY,EAAG,OAAM,KAAK,YAAY,OAAO,MAAM,SAAS;AAClF,UAAI,KAAK,cAAc,YAAY,EAAG,OAAM,KAAK,aAAa,OAAO,MAAM,SAAS;AACpF,UAAI,KAAK,aAAa,YAAY,EAAG,OAAM,KAAK,YAAY,OAAO,MAAM,SAAS;AAClF,UAAI,KAAK,gBAAgB,YAAY,EAAG,OAAM,KAAK,eAAe,OAAO,MAAM,SAAS;AAAA,IAC1F,SAAS,KAAK;AACZ,iCAAQ,wBAAwB,EAAE,QAAQ,KAAK,IAAI,OAAO,OAAO,GAAG,EAAE,CAAC;AAAA,IACzE;AAAA,EACF;AAAA;AAAA,EAGA,MAAM,gBAAiC;AACrC,QAAI,KAAK,UAAU,YAAY,EAAG,QAAO,KAAK,SAAS,MAAM;AAC7D,QAAI,KAAK,YAAY,YAAY,EAAG,QAAO,KAAK,WAAW,MAAM;AACjE,QAAI,KAAK,aAAa,YAAY,EAAG,QAAO,KAAK,YAAY,MAAM;AACnE,QAAI,KAAK,cAAc,YAAY,EAAG,QAAO,KAAK,aAAa,MAAM;AACrE,QAAI,KAAK,aAAa,YAAY,EAAG,QAAO,KAAK,YAAY,MAAM;AACnE,QAAI,KAAK,gBAAgB,YAAY,EAAG,QAAO,KAAK,eAAe,MAAM;AACzE,WAAO;AAAA,EACT;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAQA,MAAM,iBAAiB,SAAkC;AACvD,eAAW,MAAM,SAAS;AACxB,WAAK,UAAU,OAAO,EAAE;AACxB,UAAI,KAAK,YAAY,YAAY,EAAG,OAAM,KAAK,WAAW,OAAO,EAAE;AACnE,UAAI,KAAK,aAAa,YAAY,EAAG,OAAM,KAAK,YAAY,OAAO,EAAE;AACrE,UAAI,KAAK,cAAc,YAAY,EAAG,OAAM,KAAK,aAAa,OAAO,EAAE;AACvE,UAAI,KAAK,aAAa,YAAY,EAAG,OAAM,KAAK,YAAY,OAAO,EAAE;AACrE,UAAI,KAAK,gBAAgB,YAAY,EAAG,OAAM,KAAK,eAAe,OAAO,EAAE;AAAA,IAC7E;AAAA,EACF;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAcA,MAAM,SAAS,YAAwE;AACrF,QAAI,CAAC,KAAK,cAAc,KAAK,CAAC,KAAK,SAAU,QAAO;AAEpD,UAAM,UAAU,KAAK,OAAO,kBAAkB;AAC9C,UAAM,yBAAyB,CAAC,GAAG,gBAAgB;AAGnD,UAAM,kBAAkB,KAAK,GAAG,qBAAqB,sBAAsB;AAC3E,UAAM,2BAA2B;AACjC,QAAI,kBAAkB,0BAA0B;AAG9C,mBAAa,IAAI,eAAe;AAAA,IAClC;AAGA,UAAM,cAAc,IAAI;AAAA,MACtB,KAAK,gBAAgB,YAAY,IAC7B,MAAM,KAAK,eAAe,mBAAmB,IAC7C,KAAK,aAAa,YAAY,IAC9B,MAAM,KAAK,YAAY,mBAAmB,IAC1C,KAAK,cAAc,YAAY,IAC/B,MAAM,KAAK,aAAa,mBAAmB,IAC3C,KAAK,aAAa,YAAY,IAC5B,MAAM,KAAK,YAAY,mBAAmB,IAC1C,KAAK,YAAY,YAAY,IAC3B,MAAM,KAAK,WAAW,mBAAmB,IACzC,KAAK,UAAU,YAAY,IACzB,KAAK,SAAS,mBAAmB,IACjC,KAAK,GAAG,mBAAmB;AAAA,IACvC;AAGA,UAAM,YAAY;AAClB,QAAI,aAAa;AACjB,QAAI,YAAY;AAChB,QAAI,mBAAmB;AACvB,QAAI,uBAAuB;AAK3B,UAAM,eAAe,KAAK,IAAI,GAAG,kBAAkB,YAAY,IAAI;AACnE,QAAI,iBAAiB,GAAG;AACtB,UAAI,KAAK,YAAY,YAAY,EAAG,OAAM,KAAK,WAAW,KAAK;AAC/D,aAAO;AAAA,IACT;AAEA,gCAAS,6BAA6B,YAAY,WAAW,YAAY,IAAI,4CAA4C,SAAS,GAAG;AAGrI,QAAI,aAAqB,CAAC;AAC1B,QAAI,mBAAmB;AACvB,QAAI,gBAAgB;AAEpB,UAAM,gBAAgB,MAAM;AAC1B,YAAM,OAAO,KAAK,GAAG,wBAAwB,wBAAwB,WAAW,UAAU;AAC1F,oBAAc,KAAK;AACnB,mBAAa,KAAK,OAAO,OAAK,CAAC,YAAY,IAAI,EAAE,EAAE,CAAC;AACpD,yBAAmB;AACnB,UAAI,KAAK,SAAS,UAAW,iBAAgB;AAAA,IAC/C;AAEA,kBAAc;AAEd,WAAO,WAAW,SAAS,KAAM,CAAC,iBAAiB,oBAAoB,WAAW,QAAS;AAEzF,UAAI,oBAAoB,WAAW,QAAQ;AACzC,YAAI,cAAe;AACnB,sBAAc;AACd,YAAI,WAAW,WAAW,EAAG;AAAA,MAC/B;AAGA,YAAM,QAAQ,WAAW,MAAM,kBAAkB,mBAAmB,gBAAgB;AACpF,UAAI,MAAM,WAAW,EAAG;AAExB,YAAM,QAAQ,MAAM,IAAI,OAAK,oBAAoB,WAAW,CAAC,CAAC,EAAE;AAEhE,UAAI;AACF,cAAM,UAAU,MAAM,KAAK,SAAS,OAAO,EAAE,SAAS,QAAQ,WAAW,KAAK,CAAC;AAC/E,cAAM,OAAiB,QAAQ;AAC/B,cAAM,MAAM,KAAK,CAAC,MAAM,KAAK,OAAO,gBAAgB;AACpD,cAAM,OAAO,eAAe,QAAQ,IAAI;AAExC,cAAM,UAAkB,CAAC;AACzB,cAAM,eAA+B,CAAC;AACtC,cAAM,UAAkB,CAAC;AACzB,cAAM,eAA+B,CAAC;AAEtC,iBAAS,IAAI,GAAG,IAAI,MAAM,QAAQ,KAAK;AACrC,gBAAM,OAAO,MAAM,CAAC;AACpB,gBAAM,YAAY,KAAK,MAAM,IAAI,MAAM,IAAI,KAAK,GAAG;AACnD,cAAI,CAAC,KAAK,UAAU,YAAY,KAAK,CAAC,KAAK,YAAY,YAAY,KAAK,CAAC,KAAK,aAAa,YAAY,KAAK,CAAC,KAAK,cAAc,YAAY,KAAK,CAAC,KAAK,aAAa,YAAY,KAAK,CAAC,KAAK,gBAAgB,YAAY,GAAG;AACxN,iBAAK,GAAG,eAAe,KAAK,IAAI,WAAW,OAAO;AAAA,UACpD;AACA,eAAK,UAAU,OAAO,KAAK,IAAI,SAAS;AACxC,cAAI,KAAK,YAAY,YAAY,EAAG,OAAM,KAAK,WAAW,OAAO,MAAM,SAAS;AAChF,cAAI,KAAK,aAAa,YAAY,EAAG,OAAM,KAAK,YAAY,OAAO,MAAM,SAAS;AAClF,cAAI,KAAK,cAAc,YAAY,EAAG,OAAM,KAAK,aAAa,OAAO,MAAM,SAAS;AACpF,cAAI,KAAK,aAAa,YAAY,GAAG;AAAE,oBAAQ,KAAK,IAAI;AAAG,yBAAa,KAAK,SAAS;AAAA,UAAG;AACzF,cAAI,KAAK,gBAAgB,YAAY,GAAG;AAAE,oBAAQ,KAAK,IAAI;AAAG,yBAAa,KAAK,SAAS;AAAA,UAAG;AAAA,QAC9F;AAEA,YAAI,QAAQ,SAAS,EAAG,OAAM,KAAK,YAAa,WAAW,SAAS,YAAY;AAChF,YAAI,QAAQ,SAAS,EAAG,OAAM,KAAK,eAAgB,WAAW,SAAS,YAAY;AAEnF,4BAAoB,MAAM;AAC1B,qBAAa,MAAM;AACnB;AAEA,YAAI,mBAAmB,cAAc,wBAAwB,GAAG;AAC9D,6BAAmB,KAAK,IAAI,mBAAmB,GAAG,UAAU;AAC5D,sCAAS,0CAA0C,gBAAgB,EAAE;AACrE,iCAAuB;AAAA,QACzB;AAAA,MACF,SAAS,KAAK;AACZ,cAAM,SAAS,OAAO,GAAG;AACzB,cAAM,QAAQ,OAAO,SAAS,gBAAgB,KAAK,OAAO,SAAS,KAAK,KAAK,OAAO,SAAS,eAAe;AAE5G,YAAI,SAAS,mBAAmB,gBAAgB;AAC9C,6BAAmB,KAAK,IAAI,KAAK,MAAM,mBAAmB,CAAC,GAAG,cAAc;AAC5E,iCAAuB;AACvB,qCAAQ,sCAAsC,gBAAgB,kCAA6B,gBAAgB,iBAAiB;AAAA,YAC1H,WAAW;AAAA,YACX,SAAS,MAAM,MAAM,GAAG,CAAC,EAAE,IAAI,OAAK,EAAE,EAAE;AAAA,UAC1C,CAAC;AAAA,QAEH,WAAW,SAAS,oBAAoB,gBAAgB;AACtD,sCAAS,iDAAiD,cAAc,oBAAe,MAAM,MAAM,UAAU;AAAA,YAC3G,YAAY;AAAA,YACZ,cAAc,MAAM,IAAI,OAAK,GAAG,EAAE,IAAI,IAAI,EAAE,IAAI,KAAK,EAAE,QAAQ,GAAG,EAAE,MAAM,GAAG,CAAC;AAAA,UAChF,CAAC;AACD,8BAAoB,MAAM;AAC1B,uBAAa,MAAM;AACnB,iCAAuB;AAAA,QACzB,OAAO;AACL,qCAAQ,mDAAmD,gBAAgB,IAAI;AAAA,YAC7E,WAAW;AAAA,YACX,OAAO;AAAA,UACT,CAAC;AACD,8BAAoB,MAAM;AAC1B,uBAAa,MAAM;AACnB,iCAAuB;AAAA,QACzB;AAAA,MACF;AAEA,mBAAa,WAAW,YAAY;AAAA,IACtC;AAEA,QAAI,KAAK,YAAY,YAAY,EAAG,OAAM,KAAK,WAAW,KAAK;AAE/D,gCAAS,4CAAuC,SAAS,iCAAiC,gBAAgB,EAAE;AAE5G,WAAO;AAAA,EACT;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAWA,MAAM,OAAO,OAAe,OAAO,IAAqB;AACtD,QAAI,CAAC,KAAK,cAAc,KAAK,CAAC,KAAK,SAAU,QAAO,CAAC;AAErD,QAAI;AACF,YAAM,OAAO,iBAAiB,KAAK;AACnC,YAAM,SAAS,MAAM,KAAK,SAAS,MAAM,EAAE,SAAS,QAAQ,WAAW,KAAK,CAAC;AAC7E,YAAM,WAAW,eAAe,OAAO,IAAI;AAE3C,UAAI;AAEJ,UAAI,KAAK,aAAa,YAAY,GAAG;AAEnC,kBAAU,MAAM,KAAK,YAAY,OAAO,OAAO,UAAU,IAAI;AAAA,MAC/D,WAAW,KAAK,YAAY,YAAY,GAAG;AAEzC,kBAAU,MAAM,KAAK,WAAW,OAAO,OAAO,UAAU,IAAI;AAAA,MAC9D,WAAW,KAAK,cAAc,YAAY,GAAG;AAE3C,kBAAU,MAAM,KAAK,aAAa,OAAO,UAAU,IAAI;AAAA,MACzD,WAAW,KAAK,aAAa,YAAY,GAAG;AAE1C,kBAAU,MAAM,KAAK,YAAY,OAAO,UAAU,IAAI;AAAA,MACxD,WAAW,KAAK,gBAAgB,YAAY,GAAG;AAE7C,kBAAU,MAAM,KAAK,eAAe,OAAO,UAAU,IAAI;AAAA,MAC3D,WAAW,KAAK,UAAU,YAAY,GAAG;AAEvC,kBAAU,KAAK,SAAS,OAAO,UAAU,IAAI;AAAA,MAC/C,OAAO;AAEL,cAAM,gBAAgB,KAAK,GAAG,iBAAiB;AAC/C,kBAAU,cACP,IAAI,CAAC,EAAE,QAAQ,UAAU,OAAO,EAAE,QAAQ,OAAO,OAAO,UAAU,SAAS,EAAE,EAAE,EAC/E,OAAO,OAAK,EAAE,SAAS,GAAG,EAC1B,KAAK,CAAC,GAAG,MAAM,EAAE,QAAQ,EAAE,KAAK,EAChC,MAAM,GAAG,IAAI,EACb,IAAI,OAAK,EAAE,MAAM;AAAA,MACtB;AAEA,YAAM,UAAkB,CAAC;AACzB,iBAAW,UAAU,SAAS;AAC5B,cAAM,OAAO,KAAK,GAAG,QAAQ,MAAM;AACnC,YAAI,KAAM,SAAQ,KAAK,IAAI;AAAA,MAC7B;AACA,aAAO;AAAA,IACT,SAAS,KAAK;AACZ,iCAAQ,gCAAgC,EAAE,OAAO,OAAO,GAAG,EAAE,CAAC;AAC9D,aAAO,CAAC;AAAA,IACV;AAAA,EACF;AAAA;AAAA,EAGA,QAAc;AACZ,SAAK,aAAa,MAAM;AACxB,SAAK,gBAAgB,MAAM;AAAA,EAC7B;AACF;",
+  "names": []
+}

package/dist/vectors/lancedb-index.d.ts

@@ -0,0 +1,50 @@
+/**
+ * KiroGraph LanceDB Index
+ *
+ * ANN vector search backed by @lancedb/lancedb (Apache Arrow / Lance format).
+ * The database is persisted to .kirograph/lancedb/ (Lance columnar file storage).
+ *
+ * Opt-in: set config.semanticEngine = 'lancedb'
+ * Required optional dependency (not installed by default):
+ *   npm install @lancedb/lancedb
+ *
+ * Key characteristics:
+ *   - Pure JS/WASM, no native compilation required
+ *   - Columnar storage (Apache Lance format) — efficient for batch reads/writes
+ *   - ANN search via IVF-PQ or HNSW index for sub-linear query time
+ *   - Native upsert via mergeInsert (single round-trip, no delete+insert dance)
+ */
+import type { Node } from '../types';
+export declare class LanceDBIndex {
+    private readonly kirographDir;
+    private readonly dim;
+    private db;
+    private table;
+    private _available;
+    private dbPath;
+    constructor(kirographDir: string, dim?: number);
+    isAvailable(): boolean;
+    /**
+     * Load @lancedb/lancedb, open the file-persisted database, and open or create
+     * the kg_nodes table. Silent no-op when the optional dep is missing.
+     */
+    initialize(): Promise<void>;
+    /**
+     * Insert or update a node using LanceDB's native mergeInsert — a single
+     * round-trip that updates the row if node_id exists, inserts it otherwise.
+     */
+    upsert(node: Node, embedding: Float32Array): Promise<void>;
+    /**
+     * Remove a node's record from the index.
+     */
+    delete(nodeId: string): Promise<void>;
+    /**
+     * ANN vector search. Returns node IDs ordered by cosine similarity (descending).
+     */
+    search(queryVec: Float32Array, topN?: number): Promise<string[]>;
+    /** Return all node IDs currently stored in the index. */
+    getEmbeddedNodeIds(): Promise<string[]>;
+    /** Number of records currently in the index. */
+    count(): Promise<number>;
+}
+//# sourceMappingURL=lancedb-index.d.ts.map

package/dist/vectors/lancedb-index.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"lancedb-index.d.ts","sourceRoot":"","sources":["../../src/vectors/lancedb-index.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;GAeG;AAIH,OAAO,KAAK,EAAE,IAAI,EAAE,MAAM,UAAU,CAAC;AAUrC,qBAAa,YAAY;IAOrB,OAAO,CAAC,QAAQ,CAAC,YAAY;IAC7B,OAAO,CAAC,QAAQ,CAAC,GAAG;IAPtB,OAAO,CAAC,EAAE,CAAgB;IAC1B,OAAO,CAAC,KAAK,CAAa;IAC1B,OAAO,CAAC,UAAU,CAAS;IAC3B,OAAO,CAAC,MAAM,CAAS;gBAGJ,YAAY,EAAE,MAAM,EACpB,GAAG,SAAc;IAKpC,WAAW,IAAI,OAAO;IAItB;;;OAGG;IACG,UAAU,IAAI,OAAO,CAAC,IAAI,CAAC;IAsCjC;;;OAGG;IACG,MAAM,CAAC,IAAI,EAAE,IAAI,EAAE,SAAS,EAAE,YAAY,GAAG,OAAO,CAAC,IAAI,CAAC;IAqBhE;;OAEG;IACG,MAAM,CAAC,MAAM,EAAE,MAAM,GAAG,OAAO,CAAC,IAAI,CAAC;IAU3C;;OAEG;IACG,MAAM,CAAC,QAAQ,EAAE,YAAY,EAAE,IAAI,SAAK,GAAG,OAAO,CAAC,MAAM,EAAE,CAAC;IAiBlE,yDAAyD;IACnD,kBAAkB,IAAI,OAAO,CAAC,MAAM,EAAE,CAAC;IAU7C,gDAAgD;IAC1C,KAAK,IAAI,OAAO,CAAC,MAAM,CAAC;CAQ/B"}