Evaluates hybrid retrieval pipelines (BM25 + vector + reranker) end-to-end: authors ground-truth judgment sets, computes nDCG@k and MRR over fused results, measures the lift from Reciprocal Rank Fusion vs weighted fusion vs single-stage retrieval, and quantifies reranker (cross-encoder/Cohere/bge) impact. Use when a production system combines lexical and semantic retrieval and you need a numeric relevance baseline, fusion-strategy comparison, or evidence that a reranker is earning its latency cost.