Baseline vs Candidate Regression

Compare BM25, Semantic and Hybrid retrieval on the same labelled benchmark, identify query-level regressions, and turn evidence into a deterministic release decision.

BM25 → Semantic → Hybrid

Gate policy: candidate may not lose more than 2 percentage points Recall@K, 3 points MRR, 3 points nDCG@K, or any Hit Rate. RELEASE requires all guardrails plus ≥1 point composite-quality uplift. PASS means no material regression. Otherwise: REGRESSION.

Ready. Semantic and Hybrid run locally in the browser with Hugging Face Transformers.js.

Run the regression gate to compare all three retrievers.

Query-level regression analysis

Find exactly which queries improved, remained stable, or regressed

No regression run yet.

Benchmark data

Same corpus and labels are used for every retriever

Corpus CSV · doc_id,text

Evaluation CSV · query,relevant_doc_ids