ryancodrai/turbovec

turbovec is a vector search library written in Rust, with Python bindings, built for applications that need to search through large collections of embeddings, the numerical representations that AI models use to capture the meaning of text, images, or other data. It implements Google Research's TurboQuant algorithm, a technique for compressing vectors that needs no training step and no pass over your data before it starts working. The pitch behind the project is memory and speed. A ten million document collection stored as regular 32 bit floating point numbers takes about 31 gigabytes of RAM. turbovec compresses the same collection down to about 4 gigabytes while also searching faster than FAISS, a widely used vector search library from Meta. Because there is no codebook training step, you can add new vectors to the index at any time without needing to rebuild anything. Under the hood, turbovec normalizes each vector, applies a random rotation so that every coordinate follows a predictable statistical distribution, then uses that predictability to bucket each coordinate into a small number of bits using an algorithm called Lloyd Max quantization. The compressed coordinates get packed tightly into bytes, and searches are done directly against the compressed representation using hand written processor specific code for ARM and x86 chips, so there is no need to decompress vectors before comparing them. You can use turbovec from Python with a simple pip install, or from Rust with cargo. It also has ready made integrations for popular AI frameworks like LangChain, LlamaIndex, and Haystack. Because everything runs locally with no external service involved, it fits well into retrieval augmented generation, commonly called RAG, setups where data privacy, memory limits, or search speed matter, and nothing needs to leave your own machine or private network.