Hybrid search

Semantic search is great at conceptual queries ("how do I save money on infra") and bad at exact tokens (product SKUs, error codes, legal citations, names with weird spellings). Keyword search is the inverse — it nails exact matches and falls apart on paraphrase. Hybrid search runs both in parallel and merges the results, getting the strengths of each. The mechanics: index every document twice — once as a BM25 inverted index, once as a vector index. At query time, run both retrievers, get two ranked lists, then fuse them. Reciprocal Rank Fusion (RRF) is the simplest and works shockingly well — score each doc by the sum of 1/(k+rank) across both lists. Weighted score fusion (alpha * semantic_score + (1-alpha) * bm25_score) gives you a tunable knob. Most production stacks use one of these. You really see the difference on real-world queries. "How do I configure SSO for Okta" — both retrievers nail it. "What does error E1042 mean" — BM25 saves you because the embedding might match other errors. "Ways to make our docs less confusing" — semantic shines because no doc literally contains that phrase. Run only one and you'll lose 20-30% of your queries on the long tail. Most managed vector DBs (Pinecone, Qdrant, Weaviate, OpenSearch, pgvector with paradedb) ship hybrid search natively in 2026. You don't have to build it. You do have to tune the fusion weight and decide whether to rerank the merged list — usually yes.