Hybrid search
Hybrid search runs two retrievers at once: keyword search that matches exact words, and vector search that matches meaning. You blend their results into one ranked list. The catch: the two return scores on totally different scales, so how you merge them decides whether hybrid actually beats either one alone.
Why it matters
Vector search is great at "sounds like this" but fumbles literal tokens like product codes, error strings, and names. Keyword search nails those exact tokens but goes blind the moment a user phrases things differently than your docs. Lean on one alone and you get confident retrieval that quietly misses half your questions. Downstream, that shows up as thin context and weaker answers.
How it works
Send the query to both a keyword index (usually BM25) and a vector index, then fuse the two result lists. The common trick is Reciprocal Rank Fusion, which merges by rank position instead of raw scores, so the scale mismatch stops mattering. Tune the weighting toward keyword or vector depending on your data, then measure with context recall to confirm the right chunks actually show up.
A support bot gets asked "why is my card throwing err_4012?" Vector search alone drifts toward general articles about card failures and never surfaces the exact page. Add keyword search and err_4012 matches literally, so the fused result puts the right doc on top and the answer is grounded in the real error.