Dense vs sparse (BM25) vs hybrid retrieval

Sparse (BM25/TF-IDF): lexical overlap, fast, exact-match friendly, no model needed, hates paraphrase. Dense (DPR/contriever/BGE/OpenAI embeddings): semantic similarity in vector space, handles paraphrase and cross-lingual, but mismatches on rare entities and exact identifiers (model numbers, ticker symbols). Hybrid combines both, usually via score fusion (RRF = reciprocal rank fusion is the 2025 standard, not linear combination) or by training a reranker that sees both. Hybrid typically wins by 5-15% MRR/Recall over either alone, especially when queries are a mix of natural-language and keyword-style (e.g., "iPhone 15 specs"). Senior nuance: hybrid only helps if both signals complement each other; if your embedding model is great, hybrid is overhead; if your docs are highly technical with rare terms, BM25 alone can outperform dense. The right answer in 2026 is: hybrid with reranking.