Turbovec – Google's TurboQuant for vector search in Rust
Points and comments are a snapshot, not live.
Turbovec is a Rust vector index using Google's TurboQuant algorithm for memory-efficient search.
Turbovec is a Rust vector index with Python bindings, implementing Google's TurboQuant algorithm. It compresses a 10 million document corpus (31 GB as float32) into 4 GB, claiming faster search than FAISS. Features include online ingest with no training phase, SIMD-optimized search kernels, incremental save/load, and hybrid retrieval with filtered search via allowlists. Benchmarks show 3.4x speed improvement over FAISS FastScan at 4-bit on ARM, and 20% improvement on x86. The index supports removal and recall competitive with FAISS PQ.
What commenters are saying
Commenters express mixed reactions. Some praise the performance and compression, calling it 'insane' and useful for local, privacy-first search. Others criticize the README as AI-generated slop, with one calling it 'another vibe coded slop' with poor documentation. Several point to the TurboQuant paper's open review comments alleging academic misconduct and suggest RaBitQ as a better alternative. Questions arise about WASM compilation, SQLite bindings, and integrations with lanceDB and DuckDB. Some note that FAISS is no longer state-of-the-art, referencing ANN benchmarks.