turbovec publishes a Rust TurboQuant vector index with reproducible benchmarks against FAISS

Sources: turbovec repository, HN discussion
Summary: turbovec is an MIT-licensed Rust vector index with Python bindings built on Google Research's TurboQuant, a data-oblivious quantizer that needs no separate training phase, and it is a third-party implementation rather than a Google project despite the Hacker News title. Every published number states corpus, query count, k, run count, and the exact two machines, with the benchmark scripts committed and per-cell result JSON linked. Against FAISS IndexPQFastScan it reports averages of about 3.4 to 3.5 times at 4-bit and 20 to 26 percent at 2-bit on both ARM and x86, and it names the cells where plain TurboQuant trails FAISS rather than reporting only wins. The repository shows 15.3 thousand stars.
Why it matters: Online ingest with no train step, incremental crash-safe saves, and allowlist filtering applied inside the SIMD kernel are the operationally useful parts, and the stated measurement conditions make the comparison checkable rather than a launch claim.