Retrieval-Augmented Legal Question Answering Using Hybrid Semantic Search

1. A full ablation of sparse, dense, fusion-based and document-aware retrieval on the complete benchmark, scored with a character-level metric.
2. A weighted fusion scheme that repairs the failure of equal-weight reciprocal rank fusion, and a sparse backend (bm25s) that cuts BM25 scoring time by three orders of
magnitude.
3. A two-stage document-aware retriever that narrows the search to the correct document before re-ranking, the strongest method we tested.
4. A citation-grounded generation stage whose citations are scored against gold spans, showing that better retrieval yields better-attributed answers even though the model tends to under-cite.