Hybrid BM25 + vector search using Reciprocal Rank Fusion. Four embedding
tiers: built-in GloVe word vectors (20k words, 50d, always available),
API provider (Ollama/OpenAI), ONNX backend (GTE-small 384d via
onnxruntime, build tag embeddings_onnx), and Hugot/GoMLX backends
(MiniLM via pure Go or XLA, build tags embeddings_hugot/embeddings_gomlx).
Pure Go HNSW index via coder/hnsw. Pluggable Provider interface with
NewLocalProvider dispatcher. CLI flags --embeddings, --embeddings-url,
--embeddings-model on serve and bridge commands. Makefile targets for
build variants and dependency setup.
Andrey Kumanyaev
d4eba9e89c