Hybrid Search and Reranking: Getting the Best of Keywords and Meaning
Keyword search is precise but literal; semantic search understands intent but can be vague. Combining them, then reranking, is how production systems get both.
Hybrid Search and Reranking
Semantic search finds passages that mean the same thing as a question, even with different words. Keyword search (BM25 and similar) finds exact terms. Real documents need both: a lawyer searching for “clause 14.2” wants the exact reference, while an engineer asking “how do we roll back a release?” wants the runbook that never uses those words.
How hybrid search works
- Run keyword and vector retrieval side by side.
- Fuse the two ranked lists into one candidate set.
- Rerank the candidates with a cross-encoder or LLM that reads question and passage together.
- Apply filters and permissions, and return the top passages with sources.
Why reranking matters
First-stage retrieval is fast but approximate. A reranker spends more compute on a short list, which is where most of the accuracy gain comes from.
troveGEN runs this whole sequence behind a single API call. See how it works.
