RuVector ONNX Embeddings WASM

Portable embedding generation that runs anywhere WebAssembly runs

Generate text embeddings directly in browsers, Cloudflare Workers, Deno, and any WASM runtime. Built with Tract for pure Rust ONNX inference.

Features

Quick Start

Browser (ES Modules)

<script type="module">
import init, { WasmEmbedder } from 'https://unpkg.com/ruvector-onnx-embeddings-wasm/ruvector_onnx_embeddings_wasm.js';
import { createEmbedder } from 'https://unpkg.com/ruvector-onnx-embeddings-wasm/loader.js';

// Initialize WASM
await init();

// Create embedder (downloads model automatically)
const embedder = await createEmbedder('all-MiniLM-L6-v2');

// Generate embeddings
const embedding = embedder.embedOne("Hello, world!");
console.log("Dimension:", embedding.length); // 384

// Compute similarity
const sim = embedder.similarity("I love Rust", "Rust is great");
console.log("Similarity:", sim.toFixed(4)); // ~0.85
</script>

Node.js

npm install ruvector-onnx-embeddings-wasm

import { createEmbedder, similarity, embed } from 'ruvector-onnx-embeddings-wasm/loader.js';

// One-liner similarity
const score = await similarity("I love dogs", "I adore puppies");
console.log(score); // ~0.85

// One-liner embedding
const embedding = await embed("Hello world");
console.log(embedding.length); // 384

// Full control
const embedder = await createEmbedder('bge-small-en-v1.5');
const emb1 = embedder.embedOne("First text");
const emb2 = embedder.embedOne("Second text");

Cloudflare Workers

import { WasmEmbedder, WasmEmbedderConfig } from 'ruvector-onnx-embeddings-wasm';

export default {
  async fetch(request, env) {
    // Load model from R2 or KV
    const modelBytes = await env.MODELS.get('model.onnx', 'arrayBuffer');
    const tokenizerJson = await env.MODELS.get('tokenizer.json', 'text');

    const embedder = new WasmEmbedder(
      new Uint8Array(modelBytes),
      tokenizerJson
    );

    const { text } = await request.json();
    const embedding = embedder.embedOne(text);

    return Response.json({
      embedding: Array.from(embedding),
      dimension: embedding.length
    });
  }
};

Available Models

API Reference

ModelLoader

import { ModelLoader, MODELS, DEFAULT_MODEL } from './loader.js';

// List available models
console.log(ModelLoader.listModels());

// Load with progress
const loader = new ModelLoader({
  cache: true,
  onProgress: ({ percent }) => console.log(`${percent}%`)
});

const { modelBytes, tokenizerJson, config } = await loader.loadModel('all-MiniLM-L6-v2');

WasmEmbedder

class WasmEmbedder {
  constructor(modelBytes: Uint8Array, tokenizerJson: string);

  static withConfig(
    modelBytes: Uint8Array,
    tokenizerJson: string,
    config: WasmEmbedderConfig
  ): WasmEmbedder;

  embedOne(text: string): Float32Array;
  embedBatch(texts: string[]): Float32Array;
  similarity(text1: string, text2: string): number;

  dimension(): number;
  maxLength(): number;
}

WasmEmbedderConfig

class WasmEmbedderConfig {
  constructor();
  setMaxLength(length: number): WasmEmbedderConfig;
  setNormalize(normalize: boolean): WasmEmbedderConfig;
  setPooling(strategy: number): WasmEmbedderConfig;
  // 0=Mean, 1=Cls, 2=Max, 3=MeanSqrtLen, 4=LastToken
}

Utility Functions

function cosineSimilarity(a: Float32Array, b: Float32Array): number;
function normalizeL2(embedding: Float32Array): Float32Array;
function version(): string;
function simd_available(): boolean;

Pooling Strategies

Performance

Tested with all-MiniLM-L6-v2, 128 token inputs

Comparison: Native vs WASM

Use native for: servers, high throughput, GPU acceleration Use WASM for: browsers, edge, portability

Building from Source

# Install wasm-pack
cargo install wasm-pack

# Build for web
wasm-pack build --target web

# Build for Node.js
wasm-pack build --target nodejs

# Build for bundlers (webpack, vite)
wasm-pack build --target bundler

Use Cases

Semantic Search

const embedder = await createEmbedder();

// Index documents
const docs = ["Rust is fast", "Python is easy", "JavaScript runs everywhere"];
const embeddings = docs.map(d => embedder.embedOne(d));

// Search
const query = embedder.embedOne("Which language is performant?");
const scores = embeddings.map((e, i) => ({
  doc: docs[i],
  score: cosineSimilarity(query, e)
}));
scores.sort((a, b) => b.score - a.score);
console.log(scores[0]); // { doc: "Rust is fast", score: 0.82 }

Text Clustering

const texts = [
  "Machine learning is amazing",
  "Deep learning uses neural networks",
  "I love pizza",
  "Italian food is delicious"
];

const embeddings = texts.map(t => embedder.embedOne(t));
// Use k-means or hierarchical clustering on embeddings

RAG (Retrieval-Augmented Generation)

// Build knowledge base
const knowledge = [
  "RuVector is a vector database",
  "Embeddings capture semantic meaning",
  // ... more docs
];
const knowledgeEmbeddings = knowledge.map(k => embedder.embedOne(k));

// Retrieve relevant context for LLM
function getContext(query, topK = 3) {
  const queryEmb = embedder.embedOne(query);
  const scores = knowledgeEmbeddings.map((e, i) => ({
    text: knowledge[i],
    score: cosineSimilarity(queryEmb, e)
  }));
  return scores.sort((a, b) => b.score - a.score).slice(0, topK);
}

Related Packages

License

MIT License - See LICENSE for details.

Part of the RuVector ecosystem
High-performance vector operations in Rust