colsearch
A single-node, multi-vector retrieval engine built around late-interaction MaxSim scoring, running on CPU or GPU with a FastAPI server, for on-premises RAG.
colsearch is a late-interaction retrieval engine that keeps MaxSim, the ColBERT-style scoring function, as the final scorer rather than an optional reranking step, and is built so a single machine can serve it. It was renamed from voyager-index; the old import path still works with a deprecation warning until the next minor release.
It offers one contract across CPU and GPU, using Rust SIMD kernels on CPU and Triton plus fused CUDA kernels on GPU, including a 1.58-bit quantized kernel roughly 6.4 times smaller than FP16 that the project reports stays within about one point of FP16 accuracy on a standard benchmark while running several times faster than a comparison engine on the same hardware. It supports ColBERT, ColPali and ColQwen models directly, and adds database semantics such as a write-ahead log, checkpointing, crash recovery and scrolling.
It installs with pip for CPU or GPU, or runs as a Docker container exposing an OpenAPI server, with a LEMUR proxy router and FAISS shortlisting in front of the final scoring stage.