Skip to main content

Embeddings

The SDK ships local embedding generation through @huggingface/transformers. Applications can embed text in the browser, in a worker, or in Node — no API calls, no round trip, no per-token cost.

This is the /embedding subpath export. It's independent of AtomicMemorySDK; the top-level class does not use the embedding module directly (backends embed server-side). The EmbeddingGenerator exists for apps that want client-side semantic search over their own data.

The pipeline​

Text goes in, vectors come out, vectors get compared, ranked results come back. The model runs as WASM inside the browser or Node process; on first use it downloads from the Hugging Face CDN and caches in IndexedDB.

Default model​

Xenova/all-MiniLM-L6-v2, 384 dimensions, int8-quantized. This matches what atomicmemory-core uses when configured with the transformers embedding provider — so the same embeddings are comparable across client and server if you ever need to mix them.

You can override the model:

import { EmbeddingGenerator } from '@atomicmemory/atomicmemory-sdk/embedding';

const generator = new EmbeddingGenerator({
  model: 'Xenova/all-MiniLM-L6-v2',
  dimensions: 384,
  provider: 'transformers',
});

const result = await generator.generateEmbedding('Hello world');
// result.embedding is a Float32Array of length 384

Caching​

First-call latency for a new model is dominated by download (~20 MB for the default) and compile. Subsequent calls in the same session are milliseconds. Across sessions, the ONNX weights are cached in IndexedDB by transformers.js itself; the SDK does not re-download.

The cache-safety helpers in src/embedding/ guard against corrupted cache entries — they validate, repair, or re-download as needed.

WASM loading​

The default transformers.js build loads its WASM runtime from the CDN. For offline or strict-CSP environments, pin a local WASM path through the transformers.js env config before instantiating the generator. Feature detection in src/embedding/feature-flags.ts selects the appropriate execution path for the runtime (WebGPU where available, WASM otherwise).

When you need this​

  • Client-side search over local data that never leaves the browser
  • Query pre-filtering before sending to a remote backend
  • Offline demos
  • Combining embeddings from the same model as the server for consistency

If you're using AtomicMemorySDK with atomicmemory-core or Mem0, you probably do not need this module — the backend embeds for you.

Next​