Turbovec: Rust vector index fits 10M docs in 4GB, claims to outrun FAISS

Turbovec is an open-source Rust vector search index with Python bindings, built on Google Research’s TurboQuant quantizer. Its headline claim is memory efficiency: a 10-million-vector corpus that consumes 31GB as float32 compresses to roughly 4GB. TurboQuant is data-oblivious, meaning it quantizes without a separate training phase or learned codebook — so vectors are indexed the moment they’re added, with no parameter tuning and no rebuilds as the corpus grows.

The project leans on hand-written SIMD kernels (NEON SDOT/SMMLA on ARM, AVX-512 VNNI on x86, with fallbacks) that the author benchmarks as beating FAISS’s IndexPQFastScan across every tested configuration — averaging about 3.4x faster at 4-bit precision and 20-23% at 2-bit. Other features target production use: crash-safe incremental saves that persist only what changed, and search-time allowlist filtering applied inside the SIMD kernel, so selective queries skip most scoring work rather than over-fetching and discarding results. Drop-in adapters ship for LangChain, LlamaIndex, Haystack, and Agno.

For a security-minded audience, the real pitch is fully local, air-gapped retrieval: no managed service, no data leaving the machine or VPC, paired with any open-source embedding model. Worth noting the benchmarks are the author’s own (100K vectors, median of 5 runs) and are strongest at high embedding dimensions; on low-dimensional GloVe data TurboQuant trails FAISS at 2-bit until a calibration pass recovers the deficit, so recall claims deserve independent verification before betting a pipeline on them.