botirk38 / turboquant

What is turboquant?

Turboquant is a Zig library implementing Google's TurboQuant algorithm for compressing high-dimensional vectors to around 3 bits per dimension, achieving near-6x size reduction with minimal distortion. It solves the storage and compute bottlenecks in vector databases and embedding search by enabling fast approximate dot products directly on compressed data—no full decode needed. Developers get an engine-based API: initialize with dimension and seed, then encode vectors, decode when required, or compute similarity scores on the fly.

Why is it gaining traction?

Unlike generic quantizers, it delivers paper-matching quality—MSE distortion below theoretical bounds, near-perfect recall@10 in benchmarks—while leveraging Zig's SIMD for ARM64 NEON speedups on encode/decode/dot ops. The hook is unbiased inner product estimates via QJL, perfect for production ANN search without retraining. As a lightweight GitHub library project akin to library google implementations, it slots easily into custom vector pipelines.

Who should use this?

ML engineers building embedding indexes for RAG or recommendation systems, where disk/RAM limits hit hard. Vector DB contributors tweaking storage layers, or edge deployers quantizing models for mobile/Arduino-like constraints. Avoid if you're locked into Python ecosystems like library google colab—stick to this for performance-critical, low-level control.

Verdict

Promising early experiment for Zig fans needing spectro turbiquant-style compression, with solid benchmarks, plots, and MIT license—but 19 stars and 1.0% credibility score signal immaturity; docs are good but expect tweaks. Try for prototypes, not yet prime-time core infra.

(198 words)