Small-scale semantic search that ditches cgo for pure Go

What it does

kelindar/search generates semantic embeddings with GGUF BERT models through llama.cpp and performs brute-force vector similarity search, all from Go. It is built for small to medium datasets—fewer than 100,000 entries—where the overhead of a full vector database is overkill. The library handles model loading, embedding computation, and index persistence, letting you save and reload vectors across restarts.

The interesting bit

Rather than wrestling with cgo, the library uses purego to call shared C libraries directly, which neatly sidesteps cross-compilation and deployment headaches. It also ships with precompiled binaries that include optional Vulkan GPU acceleration for Windows and Linux, so you can get GPU speed without building from source.

Key highlights

• Binds to llama.cpp without cgo by calling shared libraries directly via purego.
• Supports GGUF-format BERT models for embedding generation.
• Uses brute-force search with SIMD optimizations; no approximate-nearest-neighbor indexes required.
• Includes precompiled Vulkan-enabled binaries for Windows and Linux, plus CPU-only builds.
• Embeddings indexes can be saved to disk and reloaded later.

Caveats

• Datasets beyond ~100,000 entries will hit a wall because the search is strictly brute-force.
• No support for multi-field filtering, fuzzy matching, or SQL-like query operations.
• High-dimensional LLM embeddings can tax the system unless GPU resources are available.

Verdict

A solid fit for Go services that need semantic search without operating a separate vector database or touching cgo. Look elsewhere if you are indexing at scale or need complex, filtered queries.