AI engineering · 08 of 42

Hybrid Search

Combine keyword and semantic retrieval

Scroll

Combine keyword and semantic retrieval

Keyword search and vector search fail in opposite directions. Keyword search misses the page that means the same thing in other words; vector search can rank a vaguely-similar page above an exact match, because approximate similarity is all it measures.

Hybrid search runs both and merges the results into a single list, so each covers the other's blind spot.

The practical win is recall without losing precision on exact terms. Error codes, product names, version numbers and identifiers keep working, and so do the questions asked in plain language. The merging step is where the real tuning happens, and it is worth measuring rather than guessing.

Retrieval
TWO RETRIEVERS THAT FAIL DIFFERENTLY keyword (BM25) vector search exact codes, names, identifiers same meaning, different words + 1 exact fix 2 related cause 3 background BM25 alone misses the page that never says your error code. Vector search alone can rank a vaguely-similar page above an exact match, because near-enough is what it measures. Fusing the two lists gains recall without losing exact terms.
A keyword retriever and a vector retriever running together, then fusing into one ranked list.