turbovec

Why we included this project

turbovec is worth a look when your embedding collection has outgrown what a float32 index can reasonably hold. A 10 million document corpus that takes 31 GB of RAM as float32 fits in about 4 GB here, and searches come back faster than FAISS. It gets there by quantizing vectors down to 2-4 bits per component with Google's TurboQuant algorithm, which is data-oblivious and needs no separate training phase. You add vectors and they are searchable immediately, with no parameter tuning or rebuilds as the corpus grows. The hand-written SIMD kernels (AVX-512 on x86, NEON on ARM) are competitive with FAISS's product-quantization fast-scan paths, and search-time filtering by id allowlist runs inside the kernel, so selective queries skip most of the scoring work. It also ships drop-in replacements for the in-memory vector stores in LangChain, LlamaIndex, Haystack, and Agno, so swapping it into an existing retrieval pipeline is straightforward. For a self-hosted, air-gapped search layer with incremental crash-safe persistence, it is a solid, dependency-light option.