# TurboQuant WASM

> Experimental WASM + relaxed SIMD build of TurboQuant for browsers and Node.js, compressing float32 embedding vectors ~6x with fast dot product search on compressed data.

TurboQuant WASM is an npm package that brings the TurboQuant online vector quantization algorithm to browsers and Node.js via WebAssembly with relaxed SIMD acceleration. Based on the Google Research paper "TurboQuant: Online Vector Quantization with Near-optimal Distortion Rate" (ICLR 2026), it compresses float32 embedding vectors approximately 6x (~4.5 bits/dim) without any training step. It supports fast dot product search directly on compressed data, making it ideal for streaming, real-time, and edge deployments.

- **No training required** — unlike PQ/OPQ, simply call `TurboQuant.init({ dim, seed })` and start encoding any vector immediately, with no dataset-dependent configuration.
- **~6x compression** — reduces 1.5GB float32 indexes (1M × 384-dim) to ~240MB, making large embedding indexes feasible in browsers and on mobile devices.
- **Relaxed SIMD acceleration** — uses `f32x4.relaxed_madd` FMA instructions and SIMD-vectorized QJL sign packing/unpacking for maximum throughput (Chrome 114+, Firefox 128+, Safari 18+, Node.js 20+).
- **Batch search** — `dotBatch()` issues a single WASM call for all vectors, delivering up to 83x speedup over looping `dot()` calls.
- **TypeScript API** — clean async `TurboQuant.init()` / `encode()` / `decode()` / `dot()` / `dotBatch()` / `destroy()` interface with full type definitions.
- **Golden-value tests** — byte-identical output with the reference Zig implementation, verified by MSE, bits/dim, and dot product preservation metrics.
- **Single npm package** — `npm install turboquant-wasm` with embedded WASM binary and zero runtime dependencies.
- **Live demo** — vector search, image similarity, and 3D Gaussian Splatting compression running entirely in the browser.
- **MIT licensed** — free to use, modify, and distribute; built with Zig 0.15.2 and Bun.

## Features
- Online vector quantization (no training step)
- ~6x float32 compression (~4.5 bits/dim)
- Relaxed SIMD (f32x4.relaxed_madd) acceleration
- Fast dot product on compressed vectors
- Batch dot product search (dotBatch)
- TypeScript API with async init
- Embedded WASM binary (no native deps)
- Golden-value tests vs reference Zig implementation
- Browser and Node.js support
- 3D Gaussian Splatting compression demo

## Integrations
npm, Node.js, Bun, WebAssembly, TypeScript

## Platforms
MACOS, WEB, API, DEVELOPER_SDK, CLI

## Pricing
Open Source

## Version
v0.2.11

## Links
- Website: https://teamchong.github.io/turboquant-wasm/
- Documentation: https://github.com/teamchong/turboquant-wasm#readme
- Repository: https://github.com/teamchong/turboquant-wasm
- EveryDev.ai: https://www.everydev.ai/tools/turboquant-wasm