A new open source library called Turbovec reimplements TurboQuant, a vector quantization algorithm from Google Research, as a Rust based vector search index with Python bindings, according to the project's GitHub repository. The code is released under the MIT license.
Turbovec compresses a 31 gigabyte corpus of 32-bit floating point embeddings down to 4 gigabytes using 4-bit quantization, the repository says. Google describes the underlying TurboQuant algorithm as a "data oblivious quantizer with near optimal distortion and no separate training phase," published in a peer reviewed paper, according to the project.
The author's benchmarks claim Turbovec beats the FAISS library's IndexPQFastScan index by an average of 3.4 times on search speed at 4-bit compression, with single vector insertions running 7.6 to 13.9 times faster and deletions completing in roughly half a microsecond to 1.2 microseconds, according to the repository.
The project reached the front page of Hacker News this week, drawing more than 250 points, according to Hacker News.
For anyone running a vector database at meaningful scale, a sevenfold to eightfold cut in memory footprint changes the hardware math directly: fewer or cheaper machines for the same index. The numbers are the author's own benchmarks against one FAISS configuration, so builders should verify against their own workload before swapping it in.