Turbovec is a vector index built on TurboQuant, written in Rust with Python bindings.
The tool solves the problem of memory-intensive vector search by implementing Google Research's TurboQuant algorithm, a data-oblivious quantizer that requires no training phase. It compresses vector corpora dramatically—a 31 GB float32 corpus fits in 4 GB—while maintaining search speed faster than FAISS. Vectors are indexed immediately upon insertion with no training step or parameter tuning, and the index grows without rebuilds. The implementation uses hand-written SIMD kernels optimized for both ARM (NEON SDOT/SMMLA) and x86 (AVX-512 VNNI and vpermb) architectures, with AVX2 and scalar fallbacks for broader compatibility.
Adopt this tool if you need vector search with tight memory constraints, low latency, or privacy requirements that demand local-only processing. It suits RAG systems, especially those running in resource-constrained environments or requiring air-gapped deployments. The tool offers incremental persistence through sync operations that write only changed data with single fsync calls, crash-safe at any byte boundary. Filtered search is built into the SIMD kernel itself, allowing you to restrict results to an allowlist without over-fetching or recall penalties—blocks with no allowed slots are short-circuited before scoring work begins. The project compares favorably to FAISS IndexPQFastScan across measured configurations.
Maintainers respond to new issues and pull requests within a day. Work in the issue tracker centers on bug fixes, enhancements, and documentation.
