ruvector-onnx-embeddings-wasm

name: ruvector-onnx-embeddings-wasm description: "Portable WASM embedding generation using ONNX Runtime with SIMD acceleration and parallel workers. Use when generating text embeddings in browsers without a server, running embedding models at the edge, or building offline-capable semantic search applications."

ruvector-onnx-embeddings-wasm

Portable WebAssembly embedding generation powered by ONNX Runtime. Generates text embeddings directly in browsers and Node.js with SIMD acceleration and Web Worker parallelism, requiring no server-side inference.

Quick Reference

Installation

Hub install (recommended): npx agentic-flow@latest includes this package. Standalone: npx ruvector-onnx-embeddings-wasm@latest

Node.js Usage

import {
  EmbeddingModel,
  generateEmbeddings,
  cosineSimilarity,
} from 'ruvector-onnx-embeddings-wasm';

// Load a model
const model = await EmbeddingModel.load('all-MiniLM-L6-v2');

// Generate embeddings
const embeddings = await model.embed([
  'The quick brown fox',
  'A fast auburn canine',
  'Quantum computing advances',
]);

// Compare similarity
const similarity = cosineSimilarity(embeddings[0], embeddings[1]);
console.log(`Similarity: ${similarity}`); // ~0.85 (semantically similar)

const different = cosineSimilarity(embeddings[0], embeddings[2]);
console.log(`Similarity: ${different}`); // ~0.15 (semantically different)

// Batch processing with workers
const largeResults = await generateEmbeddings(thousandTexts, {
  model: 'all-MiniLM-L6-v2',
  batchSize: 64,
  numWorkers: 4,
});

Browser Usage

<script type="module">
  import { EmbeddingModel, cosineSimilarity } from 'ruvector-onnx-embeddings-wasm';

  const model = await EmbeddingModel.load('all-MiniLM-L6-v2');
  const [vecA, vecB] = await model.embed(['Hello world', 'Hi earth']);
  const score = cosineSimilarity(vecA, vecB);
  document.getElementById('score').textContent = score.toFixed(4);
</script>

Key API

EmbeddingModel

Load and run ONNX embedding models in WASM.

const model = await EmbeddingModel.load(modelId: string, options?: LoadOptions): Promise<EmbeddingModel>

LoadOptions:

Supported models:

model.embed(texts)

Generate embeddings for an array of texts.

await model.embed(texts: string[]): Promise<Float32Array[]>

model.embedOne(text)

Generate embedding for a single text.

await model.embedOne(text: string): Promise<Float32Array>

model.dimensions

model.dimensions: number  // e.g., 384

model.dispose()

Free WASM memory and ONNX session.

model.dispose(): void

generateEmbeddings(texts, config)

High-level batch embedding with parallel Web Workers.

await generateEmbeddings(texts: string[], config: BatchConfig): Promise<Float32Array[]>

BatchConfig:

cosineSimilarity(a, b)

Compute cosine similarity between two vectors.

cosineSimilarity(a: Float32Array, b: Float32Array): number  // -1 to 1

dotProduct(a, b)

dotProduct(a: Float32Array, b: Float32Array): number

euclideanDistance(a, b)

euclideanDistance(a: Float32Array, b: Float32Array): number

References

• API Reference
• npm