turbovec
A Rust vector index with Python bindings, built on Google Research's TurboQuant quantizer, that fits large embedding corpora in a fraction of the memory float32 storage needs.
turbovec is a vector index for approximate nearest-neighbor search, written in Rust with Python bindings, built on Google Research's TurboQuant algorithm, a data-oblivious quantizer with near-optimal distortion that needs no separate training phase. The project reports that a ten-million-document corpus stored as float32 takes 31 GB of RAM, while turbovec fits the same corpus in 4 GB and searches it faster than FAISS.
Vectors are indexed as soon as they are added, with no training step, parameter tuning or rebuild needed as the corpus grows, and hand-written SIMD kernels for ARM NEON and x86 AVX-512, with AVX2 and scalar fallbacks, are reported to beat FAISS's IndexPQFastScan across the measured configurations. Incremental saves persist only what changed since the last sync, with one fsync per call, and a search can be restricted to an allowed set of ids or a slot bitmask, applied directly inside the SIMD kernel rather than by filtering results afterward.
Everything runs locally, with no managed service and no data leaving the user's machine, so it can pair with any open-source embedding model for a fully offline RAG stack.