LENS
Semantic code search.
LENS tests deterministic semantic search over codebases without embedding models. The benchmark measures precision at top-5 retrieved results versus grep and ripgrep baselines.
Iterated through six versions; v0.6 is the deterministic version after a non-determinism bug was identified and fixed via 10-run variance audit. The current P@5 of 0.250 is fully deterministic and reproducible.
Non-determinism bug in v0.1–v0.5. Earlier versions had variance across runs. Identified and fixed via 10-run variance audit. v0.6 is deterministic.
Corpus-specific. Performance measured on sealed benchmark corpus. Novel codebase performance not characterized.
No embedding models. Design choice for determinism. May limit semantic understanding compared to embedding-based approaches.