Skip to main content
Runs the golden-question sets (core + paraphrase) against each mode and reports hit@5 and tokens-per-query. Run from the repo root (the corpus paths are repo-relative). This is the gate for every retrieval feature: a mode or ranking stage only becomes the default when it beats the current default here. Current standing: BM25 default; hybrid ships opt-in after tying (not beating) it.