Description
Ten million documents as float32 vectors take 31GB of RAM. turbovec is a Rust vector index built on Google's TurboQuant that fits it in 4GB and searches faster than FAISS.
TurboQuant needs no training and has near-optimal distortion, and turbovec offers Python bindings with AVX-512 and NEON acceleration.
Compression:31GB down to 4GB.
Faster:Beats FAISS.
No training:Data-oblivious quantizer.
Python:pip install.
TurboQuant needs no training and has near-optimal distortion, and turbovec offers Python bindings with AVX-512 and NEON acceleration.
Features
Compression:31GB down to 4GB.
Faster:Beats FAISS.
No training:Data-oblivious quantizer.
Python:pip install.
