embrix

Production-ready local text embeddings using @xenova/transformers. Zero external API calls, runs entirely in Node.js.

Features

• Local Execution - No API calls, runs entirely on your machine
• Two Optimized Models - MiniLM and BGE for different use cases
• Zero Dependencies - Only @xenova/transformers as dependency
• Type-Safe - Full TypeScript support with strict typing
• Efficient - Lazy loading, singleton pattern, batch processing
• Benchmark Tools - Built-in performance measurement utilities

Installation

npm install embrix

Quick Start

import { Embedder, EmbeddingModel, cosineSimilarity } from 'embrix';

// Create an embedder
const embedder = new Embedder(EmbeddingModel.MiniLM);

// Generate a single embedding
const embedding = await embedder.embed("Hello, world!");
console.log(embedding.length); // 384

// Generate batch embeddings
const embeddings = await embedder.embedBatch([
  "Hello, world!",
  "Goodbye, world!"
]);

// Compare similarity
const hello = await embedder.embed("Hello!");
const goodbye = await embedder.embed("Goodbye!");
const similarity = cosineSimilarity(hello, goodbye);
console.log(`Similarity: ${similarity}`);

Supported Models

Model	Enum	Dimensions	Description
all-MiniLM-L6-v2	EmbeddingModel.MiniLM	384	Fast and efficient, great for most use cases
bge-small-en-v1.5	EmbeddingModel.BGE	384	High quality English embeddings from BAAI

API Reference

Embedder Class

import { Embedder, EmbeddingModel } from 'embrix';

const embedder = new Embedder(EmbeddingModel.MiniLM);

Properties

• dimension: number - Embedding dimension (384 for both models)
• modelName: string - Human-readable model name
• modelType: EmbeddingModel - The model enum value

Methods

embed(text: string, options?): Promise<Float32Array>

Generate an embedding for a single text.

const vector = await embedder.embed("Your text here");

embedBatch(texts: string[], options?): Promise<Float32Array[]>

Generate embeddings for multiple texts efficiently.

const vectors = await embedder.embedBatch(["Text 1", "Text 2", "Text 3"]);

embedWithMetadata(text: string, options?): Promise<EmbeddingResult>

Generate embedding with full metadata.

const result = await embedder.embedWithMetadata("Your text");
console.log(result.model);      // EmbeddingModel.MiniLM
console.log(result.dimension);  // 384
console.log(result.embedding);  // Float32Array

Similarity Functions

import {
  cosineSimilarity,
  dotProduct,
  euclideanDistance,
  manhattanDistance,
  findMostSimilar,
  findKMostSimilar
} from 'embrix';

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

Calculate cosine similarity between two vectors. Range: [-1, 1].

const similarity = cosineSimilarity(vector1, vector2);

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

Calculate dot product of two vectors.

const product = dotProduct(vector1, vector2);

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

Calculate Euclidean (L2) distance between two vectors.

const distance = euclideanDistance(vector1, vector2);

findMostSimilar(query: Float32Array, candidates: Float32Array[])

Find the most similar vector to a query.

const query = await embedder.embed("search query");
const docs = await embedder.embedBatch(["doc 1", "doc 2", "doc 3"]);
const best = findMostSimilar(query, docs);
console.log(`Best match: index ${best.index}, similarity ${best.similarity}`);

findKMostSimilar(query: Float32Array, candidates: Float32Array[], k: number)

Find the k most similar vectors.

const top5 = findKMostSimilar(query, docs, 5);

Model Loading

import { preloadModel, preloadAllModels, isModelLoaded, clearModelCache } from 'embrix';

// Preload a specific model
await preloadModel(EmbeddingModel.MiniLM);

// Preload all models
await preloadAllModels();

// Check if model is loaded
if (isModelLoaded(EmbeddingModel.MiniLM)) {
  // Model is ready
}

// Clear cache to free memory
clearModelCache();

Benchmark Utilities

import { runBenchmark, runAllBenchmarks, formatBenchmarkResult } from 'embrix';

// Benchmark a single model
const results = await runBenchmark(EmbeddingModel.MiniLM);
console.log(formatBenchmarkResult(results));

// Benchmark all models
const allResults = await runAllBenchmarks();

CLI Benchmark

Run benchmarks from the command line:

npm run benchmark

# Options
npm run benchmark -- --model minilm
npm run benchmark -- --batch-size 50
npm run benchmark -- --help

Experimental Benchmark Results

For comprehensive experimental benchmark results, including performance comparisons across different models and configurations, see the test-embrix-experimental repository.

Example Output

============================================================
Benchmark: all-MiniLM-L6-v2
Model: minilm
Dimension: 384
============================================================

Running cold start benchmark...
  Duration: 2345.67ms

Running warm start benchmark...
  Duration: 12.34ms
  Speedup: 190.15x faster than cold start

Running batch benchmark (100 texts)...
  Total duration: 567.89ms
  Avg per embedding: 5.68ms
  Throughput: 176.09 embeddings/sec

Architecture

embrix/
├── src/
│   ├── models.ts      # Model definitions and metadata
│   ├── loader.ts      # Lazy singleton model loader
│   ├── embedder.ts    # Core embedding class
│   ├── similarity.ts  # Vector similarity functions
│   ├── benchmark.ts   # Performance measurement utilities
│   └── index.ts       # Barrel export
├── scripts/
│   └── benchmark.ts   # CLI benchmark script
├── examples/
│   └── usage.ts       # Usage examples
├── package.json
├── tsconfig.json
└── README.md

Design Decisions

Lazy Singleton Loading

Models are loaded on-demand and cached in memory. This ensures:

• Fast subsequent calls (warm start)
• No duplicate model loads
• Memory efficiency

Float32Array Throughout

All embeddings are returned as Float32Array for:

• Consistency with the underlying tensor output
• Memory efficiency vs regular arrays
• Compatibility with WebAssembly and GPU operations

No External Dependencies

Only @xenova/transformers is required. This keeps the package:

• Lightweight
• Secure
• Easy to audit

Requirements

• Node.js >= 18.0.0
• ~500MB disk space for model cache (first run)

Citation

If you use embrix in your research, please cite:

@software{embrix2026,
  title = {embrix: Production-Ready Local Text Embeddings for Node.js},
  author = {Tahsin Özgür Koç},
  year = {2026},
  url = {https://github.com/tahsinkoc/embrix}
}

License

MIT