Learn / Search quality
What is hybrid search, and why add reranking?
Short answer
Hybrid search runs keyword search and meaning-based vector search side by side and merges the results, because each finds what the other misses. A reranker then re-orders the merged candidates by how well each one actually answers the question.
With troveGEN
troveGEN's search call is hybrid by default: keyword and meaning search run together, results are merged and reranked, and filters and permissions are applied inside both.
Two ways of matching, two kinds of mistake
Vector search finds passages that mean the same thing even when the words differ, so it handles paraphrase well. It is weaker on exact strings: product codes, drug names, clause numbers and people's names may not be represented sharply in an embedding.
Keyword search, usually scored with BM25, rewards passages that contain the question's rare words. It is precise on identifiers and weak on different wording. Used alone, each has predictable blind spots.
Merging the two lists
The scores from the two methods are on different scales, so adding them directly does not work well. A popular fix is reciprocal rank fusion, which ignores the raw scores and combines the positions: a passage ranked high in both lists rises to the top, and one that only one method found still gets a place.
What a reranker adds
The first stage is built for speed over many candidates. A reranker, often a cross-encoder model, then reads the question and each candidate passage together and scores how well that passage answers that question. It is too slow to run over a whole collection, so it is applied only to the top few dozen candidates.
This two-stage design gives you recall from the first stage and precision from the second.
Filters, permissions and diversity
Metadata filters (date, document type, department) and permission checks should be applied inside the search, not after it, so that results are not silently thinned out. Diversity methods such as maximal marginal relevance avoid returning five near-identical passages.
Key takeaways
- Keyword and vector search fail in opposite ways, so combine them.
- Merge by rank, not by raw score.
- Rerank only a short candidate list, and apply filters and permissions inside the search.
How troveGEN helps with hybrid search and reranking
troveGEN's search call combines vector and keyword retrieval, merges them by rank, optionally diversifies and re-ranks with a cross-encoder, and applies your filters and each user's permissions inside both halves. You can compare settings on your own questions with the built-in evaluation tools.
What troveGEN provides
- Vector and keyword retrieval merged by rank, with optional diversification
- Reranking with a hosted cross-encoder, a self-hosted one, or a language model
- One filter language enforced on both halves of the search
- Per-user permissions applied before ranking
- Settings you can compare side by side with the evaluation tools
Frequently asked questions
Is hybrid search always better?
Usually, and especially on collections with codes, names or technical terms. If your content is purely conversational, vector search alone can be close, but adding keyword search rarely hurts.
Does reranking slow things down?
It adds some latency because the reranker reads each candidate, which is why it is limited to a short list. For many applications the quality gain is worth it.
What is BM25?
A long-established keyword scoring formula that favours passages containing the question's rarer words, adjusted for passage length.
How does troveGEN help with hybrid search and reranking?
troveGEN's search call is hybrid by default: keyword and meaning search run together, results are merged and reranked, and filters and permissions are applied inside both. It provides: Vector and keyword retrieval merged by rank, with optional diversification; Reranking with a hosted cross-encoder, a self-hosted one, or a language model; One filter language enforced on both halves of the search; Per-user permissions applied before ranking; Settings you can compare side by side with the evaluation tools.