#1. The Failure of Pure Dense Retrieval

Pure dense vector search struggles with out-of-vocabulary technical acronyms, exact product SKUs, part numbers, and legal citations. Conversely, pure keyword search (BM25) fails on synonymy and semantic paraphrase. High-reliability RAG demands a hybrid approach.

Reciprocal Rank Fusion (RRF)

RRF combines ranked candidate lists from heterogeneous retrieval algorithms without requiring calibrated score normalization:

where $k$ is a smoothing constant (typically 60) and $r_m(d)$ is the rank of document $d$ in system $m$.

---

#2. Two-Stage Retrieval with Cross-Encoders

• Stage 1 (Candidate Generation): Fast approximate search over vector and BM25 indices fetches top 50 to 100 candidate chunks.
• Stage 2 (Cross-Encoder Reranking): A deep transformer (e.g. bge-reranker-large) simultaneously evaluates the concatenated query-document pair, capturing cross-attention token interactions to filter the top 5 to 10 highest-quality passages.