LANGUAGES Signal 506
Turbovec compresses 31 GB float32 vectors to 4 GB, searches faster than FAISS
Turbovec is a Rust vector index with Python bindings implementing Google's TurboQuant algorithm, compressing a 10 million document corpus from 31 GB to 4 GB while outperforming FAISS IndexPQFastScan searches.
Engineers building vector search or RAG systems get near-8× memory compression and faster queries without a separate training phase, parameter tuning, or index rebuilds. The pure-local, incremental-save design suits privacy-sensitive and latency-critical deployments where managed services are unacceptable.
Written by elseif from the cluster below · every claim links back to a sourceThe three things worth knowing
Turbovec's data-oblivious TurboQuant quantizer requires no training phase, vectors are indexed on ingest with no parameter tuning or rebuilds as the corpus grows.
Hand-written SIMD kernels on ARM (NEON SDOT/SMMLA) and x86 (AVX-512 VNNI, vpermb) beat FAISS IndexPQFastScan in every measured config, averaging 3.4× faster at 4-bit and 23% faster at 2-bit.
Incremental sync() persists only changes since the last save with one fsync per call, and filtered search short-circuits disallowed 32-vector blocks inside the SIMD kernel to avoid over-fetching.
THE CLUSTER