How does LanceDB implement hybrid search using reranking?

Hybrid search in LanceDB runs a vector query and a full-text query independently against the same table, then hands both ranked result lists to a reranker component whose job is to merge them into one final, unified ranking rather than simply picking one list over the other.

Because vector distance scores and BM25 relevance scores aren't on the same scale or meaningfully comparable directly, a naive merge by raw score would be misleading; this is why LanceDB's default reranker uses Reciprocal Rank Fusion, which combines results based on each item's rank position in each list rather than its raw score, sidestepping the scale-comparison problem entirely.

LanceDB also supports pluggable rerankers beyond RRF — including cross-encoder-based rerankers that re-score candidates using a more expensive but more accurate model — for cases where the extra latency cost of a heavier reranking step is worth the improved relevance of the final results.