Hybrid Search: Keyword + Vector

Hybrid

Part of: Vector Databases & Production Search

A user searches for error code "E-4021". Your beautiful semantic search returns documents about general error handling , relevant in meaning, useless in fact, because it never anchored on that exact string. Meanwhile a dumb keyword search would have nailed it instantly. The fix isn't to pick a side. It's hybrid search : run both and combine them. What it is Hybrid search blends two retrieval methods: - Keyword (lexical) search matches exact terms, using classic ranking like BM25 . It excels at names, codes, IDs, and rare exact strings. - Vector (semantic) search matches meaning. It excels at paraphrases, synonyms, and fuzzy intent ("how do I stop getting charged?" finding a doc titled "Cancel your subscription"). Each is strong exactly where the other is weak. Hybrid search runs both and fuses their results into one ranked list. How it works The classic fusion method is Reciprocal Rank Fusion (RRF) . Instead of trying to compare a BM25 score to a cosine score (which live on totally different scales), RRF uses only each result's rank in each list: A document ranked near the top of either list gets a strong boost; a document ranked highly in both rises to the top. The constant k (oft

Challenge: Weighted Reciprocal Rank Fusion