@wrongstack/vector-memory 1.0.7 → 1.0.9

package/dist/index.js

@@ -378,8 +378,13 @@ function createSageSurfaceSyncSource(sage, opts = {}) {
       const requested = limit === void 0 ? maxTotal : clamp(limit, 1, maxTotal === Number.POSITIVE_INFINITY ? limit : maxTotal);
       const memories = [];
       let cursor;
+      const seenCursors = /* @__PURE__ */ new Set();
       let noProgressPages = 0;
       while (memories.length < requested) {
+        if (cursor !== void 0) {
+          if (seenCursors.has(cursor)) break;
+          seenCursors.add(cursor);
+        }
         const remaining = requested - memories.length;
         const page = await sage.listSagePage({
           statuses: ["active"],
@@ -402,6 +407,7 @@ function createSageSurfaceSyncSource(sage, opts = {}) {
           });
         }
         if (!page.nextCursor) break;
+        if (seenCursors.has(page.nextCursor)) break;
         if (rows.length === 0) {
           noProgressPages++;
           if (noProgressPages > 3) break;
@@ -929,9 +935,9 @@ var VectorMemoryStore = class _VectorMemoryStore {
    * Embed `text`, hitting the provider-level cache first. Cache miss falls
    * through to the configured provider and writes the result back. Returns
    * `undefined` when the provider fails — the caller can persist the entry
-   * without a vector (fail-open).
+   * without a vector (fail-open) — unless `strict`, which throws instead.
    */
-  async embedWithCache(text) {
+  async embedWithCache(text, strict = false) {
     const now = (/* @__PURE__ */ new Date()).toISOString();
     const cached = this.cachedVector(text, now);
     if (cached) return cached;
@@ -940,7 +946,13 @@ var VectorMemoryStore = class _VectorMemoryStore {
       const vec = result[0];
       if (vec) this.cacheVector(text, vec, now);
       return vec;
-    } catch {
+    } catch (err) {
+      if (strict) {
+        throw new VectorMemoryProviderUnavailableError(
+          `Embedding provider "${this.provider.id}" failed: ${err instanceof Error ? err.message : String(err)}`,
+          err
+        );
+      }
       return void 0;
     }
   }
@@ -1097,8 +1109,16 @@ var VectorMemoryStore = class _VectorMemoryStore {
     const threshold = opts.threshold ?? 0;
     const includeVectors = opts.includeVectors === true;
     if (typeof query !== "string" || query.trim().length === 0) return [];
-    const queryVec = await this.embedWithCache(query);
-    if (!queryVec || queryVec.length === 0) return [];
+    const strict = opts.failOnEmbeddingError === true;
+    const queryVec = await this.embedWithCache(query, strict);
+    if (!queryVec || queryVec.length === 0) {
+      if (strict) {
+        throw new VectorMemoryProviderUnavailableError(
+          `Embedding provider "${this.provider.id}" returned no vector for the query.`
+        );
+      }
+      return [];
+    }
     const providerId = this.provider.id;
     const dimensions = this.provider.dimensions;
     const filters = ["v.provider_id = ?", "v.dimensions = ?"];
@@ -1421,7 +1441,7 @@ function vectorMemorySearchTool(store) {
   return {
     name: "vector_memory_search",
     category: "Memory",
-    description: "Semantic search over the vector memory store. Embeds the query with the active provider and returns the top-k entries ranked by cosine similarity. Returns an empty list when the embedding provider is unavailable \u2014 callers should fall back to lexical search.",
+    description: "Semantic search over the vector memory store. Embeds the query with the active provider and returns the top-k entries ranked by cosine similarity. Fails when the embedding provider is unavailable \u2014 fall back to lexical search (sage `memory_search`) in that case.",
     usageHint: "Use when you want results ranked by meaning. Pairs well with sage `memory_search` for keyword precision.",
     permission: "auto",
     mutating: false,
@@ -1464,7 +1484,9 @@ function vectorMemorySearchTool(store) {
         limit: input.limit !== void 0 ? input.limit : void 0,
         threshold: input.threshold !== void 0 ? input.threshold : void 0,
         scope: input.scope,
-        kind: input.kind
+        kind: input.kind,
+        // An embedding outage must fail the call, not look like "no matches".
+        failOnEmbeddingError: true
       });
       return {
         hits: hits.map((h) => ({
@@ -1522,7 +1544,12 @@ function vectorMemoryForgetTool(store) {
       required: ["id"],
       additionalProperties: false
     },
-    execute: async (input) => ({ removed: await store.forget(input.id) })
+    execute: async (input) => {
+      if (!await store.forget(input.id)) {
+        throw new Error(`vector_memory_forget: no entry with id "${input.id}".`);
+      }
+      return { removed: true };
+    }
   };
 }
 
@@ -1547,7 +1574,8 @@ var TransformersEmbeddingProvider = class {
     this.cacheDir = opts.cacheDir;
     this.dtype = opts.dtype ?? DEFAULT_VECTOR_DTYPE;
     this.device = opts.device ?? "cpu";
-    this.batchSize = opts.batchSize ?? 16;
+    const batchSize = opts.batchSize ?? 16;
+    this.batchSize = Number.isFinite(batchSize) && batchSize >= 1 ? Math.floor(batchSize) : 16;
     this.maxChars = opts.maxChars ?? 2048;
     this.allowRemote = opts.allowRemoteModels ?? true;
     this.dimensions = DEFAULT_VECTOR_DIMENSIONS;
@@ -1595,7 +1623,15 @@ var TransformersEmbeddingProvider = class {
       if (Array.isArray(nested) && Array.isArray(nested[0])) {
         return nested.map((row) => Float32Array.from(row));
       }
-      return [Float32Array.from(nested)];
+      const flat2 = nested;
+      if (batchSize > 1 && flat2.length > 0 && flat2.length % batchSize === 0) {
+        const dimensions = flat2.length / batchSize;
+        return Array.from(
+          { length: batchSize },
+          (_, index) => Float32Array.from(flat2.slice(index * dimensions, (index + 1) * dimensions))
+        );
+      }
+      return [Float32Array.from(flat2)];
     }
     const flat = out.data;
     if (flat instanceof Float32Array) {

package/dist/store.d.ts

@@ -37,7 +37,7 @@ export declare class VectorMemoryStore {
      * Embed `text`, hitting the provider-level cache first. Cache miss falls
      * through to the configured provider and writes the result back. Returns
      * `undefined` when the provider fails — the caller can persist the entry
-     * without a vector (fail-open).
+     * without a vector (fail-open) — unless `strict`, which throws instead.
      */
     private embedWithCache;
     /**

package/dist/types.d.ts

@@ -55,6 +55,13 @@ export interface VectorSearchOptions {
      * build the pairwise-similarity heatmap.
      */
     includeVectors?: boolean | undefined;
+    /**
+     * When true, a failed query embedding THROWS
+     * `VectorMemoryProviderUnavailableError` instead of returning `[]`. The
+     * default (fail-open) suits fusion callers; the agent tool sets it so an
+     * outage is not reported as "no matches".
+     */
+    failOnEmbeddingError?: boolean | undefined;
 }
 export interface VectorSearchHit {
     entry: VectorEntry;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@wrongstack/vector-memory",
-  "version": "1.0.7",
+  "version": "1.0.9",
   "license": "MIT",
   "description": "WrongStack Vector Memory — an additional vector-search memory store powered by @huggingface/transformers (local ONNX embeddings), alongside the SAGE lexical memory system.",
   "repository": {
@@ -27,14 +27,14 @@
     "README.md"
   ],
   "dependencies": {
-    "@wrongstack/core": "1.0.7",
-    "@wrongstack/persistence": "1.0.7",
-    "@wrongstack/sage": "1.0.7"
+    "@wrongstack/core": "1.0.9",
+    "@wrongstack/persistence": "1.0.9",
+    "@wrongstack/sage": "1.0.9"
   },
   "devDependencies": {
-    "@types/node": "^26.2.0",
+    "@types/node": "^26.5.1",
     "typescript": "^7.0.2",
-    "vitest": "^4.1.11"
+    "vitest": "^5.0.0"
   },
   "publishConfig": {
     "access": "public",