Two-stage RAG retrieval: hybrid search to top-50, then cross-encoder rerank to top-5

AppScale·high signal

Retrieve a broad candidate set (N=50–100) with hybrid dense+BM25 search, then re-score with a query-aware cross-encoder and keep only the top 3–5 chunks for the LLM. On the BRIGHT Biology benchmark this pushed nDCG@10 from 0.13 to 0.40 (3x) by reordering the same candidates — reranking is the highest-ROI single change for an existing RAG pipeline.