stellavault 0.7.2 → 0.7.3

package/README.md

@@ -232,9 +232,10 @@ node tests/stress.mjs 500     # Test with 500 synthetic documents
 ```
 
 Key optimizations:
+- **HNSW graph building** — sqlite-vec KNN for 200+ docs (O(n·K·log n) vs O(n²))
 - Pre-normalized vectors: cosine similarity → single dot product
 - Batched embedding loading (500/batch, prevents RAM overflow)
-- Upper-triangle graph building (50% fewer comparisons)
+- Upper-triangle brute-force for small vaults (< 200 docs)
 - O(n) K-Means centroid updates with typed arrays
 
 ---

package/dist/stellavault.js

@@ -645,33 +645,50 @@ async function buildGraphData(store, options = {}) {
   ]);
   const edges = [];
   const edgeCounts = /* @__PURE__ */ new Map();
-  const normalizedVecs = /* @__PURE__ */ new Map();
-  for (const doc of docs) {
-    const vec = embeddings.get(doc.id);
-    if (!vec)
-      continue;
-    normalizedVecs.set(doc.id, normalizeVector([...vec]));
-  }
-  const docIds = [...normalizedVecs.keys()];
-  const vecArray = docIds.map((id) => normalizedVecs.get(id));
-  const n = docIds.length;
-  const neighbors = Array.from({ length: n }, () => []);
-  for (let i = 0; i < n; i++) {
-    for (let j = i + 1; j < n; j++) {
-      const sim = dotProduct(vecArray[i], vecArray[j]);
-      if (sim >= edgeThreshold) {
-        neighbors[i].push({ peer: j, sim });
-        neighbors[j].push({ peer: i, sim });
+  const USE_HNSW_THRESHOLD = 200;
+  const docsWithVecs = docs.filter((d) => embeddings.has(d.id));
+  if (docsWithVecs.length > USE_HNSW_THRESHOLD) {
+    for (const doc of docsWithVecs) {
+      const vec = embeddings.get(doc.id);
+      const neighbors = await store.findDocumentNeighbors(vec, maxEdgesPerNode + 1);
+      for (const { documentId: targetId, similarity } of neighbors) {
+        if (targetId === doc.id)
+          continue;
+        if (similarity < edgeThreshold)
+          continue;
+        const edgeKey = [doc.id, targetId].sort().join(":");
+        if (!edgeCounts.has(edgeKey)) {
+          edges.push({ source: doc.id, target: targetId, weight: similarity });
+          edgeCounts.set(edgeKey, 1);
+        }
       }
     }
-  }
-  for (let i = 0; i < n; i++) {
-    neighbors[i].sort((a, b) => b.sim - a.sim);
-    for (const { peer: j, sim } of neighbors[i].slice(0, maxEdgesPerNode)) {
-      const edgeKey = i < j ? `${i}:${j}` : `${j}:${i}`;
-      if (!edgeCounts.has(edgeKey)) {
-        edges.push({ source: docIds[i], target: docIds[j], weight: sim });
-        edgeCounts.set(edgeKey, 1);
+  } else {
+    const normalizedVecs = /* @__PURE__ */ new Map();
+    for (const doc of docsWithVecs) {
+      normalizedVecs.set(doc.id, normalizeVector([...embeddings.get(doc.id)]));
+    }
+    const docIds = [...normalizedVecs.keys()];
+    const vecArray = docIds.map((id) => normalizedVecs.get(id));
+    const n = docIds.length;
+    const neighbors = Array.from({ length: n }, () => []);
+    for (let i = 0; i < n; i++) {
+      for (let j = i + 1; j < n; j++) {
+        const sim = dotProduct(vecArray[i], vecArray[j]);
+        if (sim >= edgeThreshold) {
+          neighbors[i].push({ peer: j, sim });
+          neighbors[j].push({ peer: i, sim });
+        }
+      }
+    }
+    for (let i = 0; i < n; i++) {
+      neighbors[i].sort((a, b) => b.sim - a.sim);
+      for (const { peer: j, sim } of neighbors[i].slice(0, maxEdgesPerNode)) {
+        const edgeKey = i < j ? `${i}:${j}` : `${j}:${i}`;
+        if (!edgeCounts.has(edgeKey)) {
+          edges.push({ source: docIds[i], target: docIds[j], weight: sim });
+          edgeCounts.set(edgeKey, 1);
+        }
       }
     }
   }
@@ -705,16 +722,16 @@ async function buildGraphData(store, options = {}) {
       nodeCount: folderCounts.get(i) ?? 0
     }));
   } else {
-    const docIds2 = docs.filter((d) => embeddings.has(d.id)).map((d) => d.id);
-    const vectors = docIds2.map((id) => embeddings.get(id));
-    const k = Math.min(Math.max(5, Math.round(Math.sqrt(docIds2.length / 5))), 10);
+    const docIds = docs.filter((d) => embeddings.has(d.id)).map((d) => d.id);
+    const vectors = docIds.map((id) => embeddings.get(id));
+    const k = Math.min(Math.max(5, Math.round(Math.sqrt(docIds.length / 5))), 10);
     const assignments = kMeans(vectors, k);
     const clusterDocInfos = /* @__PURE__ */ new Map();
-    for (let i = 0; i < docIds2.length; i++) {
+    for (let i = 0; i < docIds.length; i++) {
       const cId = assignments[i];
       if (!clusterDocInfos.has(cId))
         clusterDocInfos.set(cId, []);
-      const doc = docs.find((d) => d.id === docIds2[i]);
+      const doc = docs.find((d) => d.id === docIds[i]);
       if (doc)
         clusterDocInfos.get(cId).push({ id: doc.id, title: doc.title });
     }
@@ -735,8 +752,8 @@ async function buildGraphData(store, options = {}) {
       });
     }
     assignmentMap = /* @__PURE__ */ new Map();
-    for (let i = 0; i < docIds2.length; i++) {
-      assignmentMap.set(docIds2[i], assignments[i]);
+    for (let i = 0; i < docIds.length; i++) {
+      assignmentMap.set(docIds[i], assignments[i]);
     }
   }
   const connectionCounts = /* @__PURE__ */ new Map();
@@ -3258,6 +3275,21 @@ function createSqliteVecStore(dbPath, dimensions = 384) {
       }
       return result;
     },
+    async findDocumentNeighbors(embedding, limit) {
+      const rows = db.prepare(`
+        SELECT c.document_id, MIN(ce.distance) as distance
+        FROM chunk_embeddings ce
+        JOIN chunks c ON c.id = ce.chunk_id
+        WHERE ce.embedding MATCH ?
+        GROUP BY c.document_id
+        ORDER BY distance
+        LIMIT ?
+      `).all(float32Buffer(embedding), limit * 2);
+      return rows.slice(0, limit).map((r) => ({
+        documentId: r.document_id,
+        similarity: 1 / (1 + r.distance)
+      }));
+    },
     async close() {
       db.close();
     },
@@ -4746,7 +4778,7 @@ function createMcpServer(options) {
   const askTool = createAskTool(searchEngine, vaultPath);
   const generateDraftTool = createGenerateDraftTool(searchEngine, vaultPath);
   const agenticTools = embedder ? createAgenticGraphTools(store, embedder, vaultPath) : [];
-  const server = new Server({ name: "stellavault", version: "0.7.2" }, { capabilities: { tools: {} } });
+  const server = new Server({ name: "stellavault", version: "0.7.3" }, { capabilities: { tools: {} } });
   server.setRequestHandler(ListToolsRequestSchema, async () => ({
     tools: [
       searchToolDef,
@@ -6902,6 +6934,7 @@ var CREDITS_FILE = join19(homedir11(), ".stellavault", "federation", "credits.js
 
 // packages/core/dist/index.js
 init_retry();
+init_math();
 init_indexer();
 function createKnowledgeHub(config) {
   const embedder = createLocalEmbedder(config.embedding.localModel);
@@ -9647,7 +9680,7 @@ if (nodeVersion < 20) {
   process.exit(1);
 }
 var program = new Command();
-var SV_VERSION = true ? "0.7.2" : "0.0.0-dev";
+var SV_VERSION = true ? "0.7.3" : "0.0.0-dev";
 program.name("stellavault").description("Stellavault \u2014 Self-compiling knowledge base for your Obsidian vault").version(SV_VERSION).option("--json", "Output in JSON format (for scripting)").option("--quiet", "Suppress non-essential output");
 program.command("init").description("Interactive setup wizard \u2014 get started in 3 minutes").action(initCommand);
 program.command("doctor").description("Diagnose setup issues (config, vault, DB, model, Node version)").action(doctorCommand);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "stellavault",
-  "version": "0.7.2",
+  "version": "0.7.3",
   "description": "Drop anything. It compiles itself into knowledge. Claude remembers everything you know. Local-first MCP server, vault files never modified.",
   "repository": {
     "type": "git",