Repository navigation
Search quality infrastructure: eval framework, fuzzy fallback, embeddings layer, fine-tuning.
Added
evalcommand: measure search quality with Precision@1, MRR, Recall@5--explain/-e: per-token hit counts, zero-match detection, quality signal--grouped/-g: directory-clustered results with--scopehints--scope/-s: SQL LIKE pre-filter for path-prefix search- Fuzzy fallback: difflib Levenshtein correction on zero results (config:
fuzzy: true) - Embeddings layer: opt-in hybrid BM25 + cosine rerank via
embeddings.enabled traincommand: TSDAE fine-tune any sentence-transformers model on codebaserecommend-modelcommand: suggest best model by codebase size- Scope-granularity eval:
file,scope,exactassert modes with line overlap - Incremental embed:
--incrementalonly re-embeds changed files
Changed
- Query semantics: AND-then-OR → pure OR + token-coverage scoring
- Default excludes: added
examples/*,docs/*,tutorials/*,benchmarks/* - JSON output:
{"results": [...]}wrapper with optionalexplain/groupedkeys
Fixed
- Flat BM25 when metadata absent — path weighting provides differentiation
- No path differentiation in ranking — filename, dir component weighting
- AND-then-OR precision hole — multi-token queries no longer fail or return noise