turbo-quant
Experimental Rust codecs for vector quantization, including PolarQuant, TurboQuant and QJL, producing derived compressed representations for semantic-search candidate generation and KV-cache compression experiments.
turbo-quant implements several derived vector-compression codecs in Rust: a seeded polar-coordinate quantizer, a version that adds a QJL residual sketch for approximate inner-product estimates, and a seeded sign-projection sketch, along with packed wire formats, candidate search, exact-rerank receipts and a KV-cache shadow-compression mode. The library does not replace a caller's exact vectors; instead it produces deterministic derived representations that can lower candidate-generation cost, alongside metadata that makes the approximate stage explicit so an exact fallback or rerank step stays part of the pipeline.
The crate targets Rust 1.75 on the 2021 edition, includes build-time C kernels for its hot paths in its published package, and states no universal quality or performance guarantee, since results depend on dimension, bits, seed, rotation and the caller's own candidate and rerank policy. The associated methods were presented at ICLR 2026.