Failure-Aware Metric Realignment for Post-Hoc Dense Retrieval

Legal retrieval method search.

Local-first retrieval research system for comparing dense retrieval, BM25, sparse-dense fusion, legal query rewriting, HyDE expansion, and cross-encoder reranking across legal and contract retrieval datasets.

Saved closure evidence includes CUAD QA, Legal RAG Bench, and MTEB Bar Exam QA result artifacts. The public summary is available while the full interactive research dashboard remains restricted until the research package is ready for release.

• Mark-2 closure: 98 completed terminal rows across three datasets.
• CUAD pivot: reranking improved Recall@5 from 0.015000 raw dense to 0.035000.
• MTEB Bar Exam QA: BM25+dense convex blend reached 0.461538 Recall@5 with a low runtime cost.