bones-search

Hybrid search engine for the bones issue tracker, combining lexical BM25, semantic embeddings, and structural graph proximity via reciprocal rank fusion (RRF).

What this crate provides

• Hybrid search: fuses three independent ranking signals with RRF
  • Lexical (FTS5/BM25): stemming, prefix search, boolean operators
  • Semantic (optional): ONNX embedding model via ort, KNN over stored vectors
  • Structural: graph proximity to lexical seed items via dependency edges
• Semantic model: loads an ONNX sentence-transformer model; gracefully degrades to lexical+structural when unavailable
• Structural similarity: shared label/parent/dependency scoring between items
• Duplicate detection: multi-signal fusion used by bn create to surface near-duplicates

Features

• semantic-ort — enable semantic search via ONNX Runtime (requires ort and tokenizers)
• bundled-model — embed a default model at build time (implies semantic-ort)

Usage

This crate is an internal dependency of bones-cli. See the bones repository for the full project.