search

Why we included this project

Semantic search in Go usually means either calling out to a separate vector database or wiring up a client library and hoping the pieces fit. This library takes a different route: it bundles embedding generation and vector lookup into a single embeddable package. It loads GGUF BERT models through llama.cpp, computes embeddings in-process, and stores them in a simple index you can persist to disk and query with brute-force similarity. The design is honest about its limits, targeting datasets under roughly 100,000 entries, where SIMD-optimized exhaustive search stays fast enough that you can skip the operational overhead of a dedicated vector store. It also calls llama.cpp through purego instead of cgo, which keeps cross-compilation and deployment of Go binaries simple, and Vulkan GPU acceleration is available on Windows and Linux when you need the extra speed.