fix(docker): use glibc base image so local embeddings load

.env.example

@@ -35,7 +35,8 @@ VOYAGE_API_KEY=
 # EMBEDDING_DIMENSIONS=512
 #
 # Local embeddings: runs a sentence-transformers model in-process via
-# transformers.js (WASM), no API key. Downloads the model once to the cache dir.
+# transformers.js (native onnxruntime), no API key. Downloads the model once to
+# the cache dir. Requires the glibc-based image (the Dockerfile uses node:slim).
 # EMBEDDING_PROVIDER=local
 # EMBEDDING_MODEL=Xenova/all-MiniLM-L6-v2
 # EMBEDDING_DIMENSIONS=384

Dockerfile

@@ -1,5 +1,5 @@
 # Stage 1: Build
-FROM node:22-alpine AS builder
+FROM node:22-slim AS builder
 WORKDIR /app
 COPY package.json package-lock.json ./
 RUN npm ci --ignore-scripts
@@ -8,7 +8,7 @@ COPY src/ src/
 RUN npx nest build
 
 # Stage 2: Production
-FROM node:22-alpine AS runner
+FROM node:22-slim AS runner
 WORKDIR /app
 ENV NODE_ENV=production
 

docker-compose.yml

@@ -18,7 +18,13 @@ services:
       redis:
         condition: service_healthy
     healthcheck:
-      test: ['CMD', 'wget', '--spider', '-q', 'http://localhost:3000/health/ready']
+      test:
+        [
+          'CMD',
+          'node',
+          '-e',
+          "fetch('http://localhost:3000/health/ready').then((r) => process.exit(r.ok ? 0 : 1)).catch(() => process.exit(1))",
+        ]
       interval: 10s
       timeout: 5s
       retries: 5

src/memory/infrastructure/adapters/local-embedding.adapter.spec.ts

@@ -25,9 +25,7 @@ describe('LocalEmbeddingAdapter', () => {
     const result = await new LocalEmbeddingAdapter(buildConfig()).embedBatch(['hello']);
 
     expect(result).toEqual([[0.1, 0.2, 0.3]]);
-    expect(pipelineMock).toHaveBeenCalledWith('feature-extraction', 'Xenova/all-MiniLM-L6-v2', {
-      device: 'wasm',
-    });
+    expect(pipelineMock).toHaveBeenCalledWith('feature-extraction', 'Xenova/all-MiniLM-L6-v2');
     expect(extractor).toHaveBeenCalledWith(['hello'], { pooling: 'mean', normalize: true });
   });
 
@@ -42,9 +40,7 @@ describe('LocalEmbeddingAdapter', () => {
     await adapter.embed('b');
 
     expect(pipelineMock).toHaveBeenCalledTimes(1);
-    expect(pipelineMock).toHaveBeenCalledWith('feature-extraction', 'Xenova/bge-small-en', {
-      device: 'wasm',
-    });
+    expect(pipelineMock).toHaveBeenCalledWith('feature-extraction', 'Xenova/bge-small-en');
   });
 
   it('returns the single vector from embed()', async () => {

src/memory/infrastructure/adapters/local-embedding.adapter.ts

@@ -6,14 +6,6 @@ import { AppConfig } from '../../../config/app.config.js';
 
 const DEFAULT_LOCAL_MODEL = 'Xenova/all-MiniLM-L6-v2';
 
-// The runner image is Alpine (musl), where the native onnxruntime-node binary
-// will not load. Force the WASM backend and a single thread so embeddings work
-// without SharedArrayBuffer.
-const wasmBackend = env.backends?.onnx?.wasm;
-if (wasmBackend) {
-  wasmBackend.numThreads = 1;
-}
-
 @Injectable()
 export class LocalEmbeddingAdapter extends EmbeddingProviderPort {
   private readonly logger = new Logger(LocalEmbeddingAdapter.name);
@@ -47,7 +39,7 @@ export class LocalEmbeddingAdapter extends EmbeddingProviderPort {
 
   private loadExtractor(): Promise<FeatureExtractionPipeline> {
     if (!this.extractor) {
-      this.extractor = pipeline('feature-extraction', this.model, { device: 'wasm' });
+      this.extractor = pipeline('feature-extraction', this.model);
     }
     return this.extractor;
   }