Hybrid search

Hybrid search combines two retrievers with opposite strengths. Vector search finds paraphrases and conceptual matches; keyword search nails exact identifiers, names, and rare terms. Running both and merging the ranked lists covers far more real queries than either alone, which is why it is a common default for production retrieval.

The merge step is the interesting part. The two retrievers produce scores on completely different scales, so simply adding them doesn't work well. Reciprocal rank fusion sidesteps the problem by combining positions instead of scores. Score normalization with weights can do slightly better when tuned on labeled data but is more fragile.

In Postgres, hybrid search can be a single query: a full-text CTE and a vector CTE, each returning its top candidates with a rank, merged with a fusion formula, all filtered to the user's workspace. Many vector databases and search services also offer hybrid queries natively.

Measure before and after. Hybrid usually helps most on identifier-heavy and short queries and should not hurt paraphrased ones; if it does, adjust candidate counts or weights. A reranker on top of the merged list is the common next step.