harshaA document can be relevant because it contains the exact term in a question. It can also be relevant because it describes the same idea using different words. Neither signal covers every case.
That is the attraction of hybrid retrieval: preserve more than one way for useful evidence to become a candidate.
Keep the strengths distinct
An illustrative query for an internal reference number needs exact matching to matter. A query phrased as “how do I restore access after leaving a team?” may need passages that use different language. Dense similarity and lexical retrieval approach those cases differently.
Combining candidates does not mean their raw scores are directly comparable. The ranking design needs a deliberate fusion strategy or another documented method. Otherwise, one score range can dominate for numerical reasons rather than relevance.
The pgvector project discusses combining vector search with PostgreSQL full-text search and points to rank-fusion approaches. Its hybrid-search section is a useful starting reference.
Evaluate the hard queries
I would build a small test set with exact identifiers, abbreviations, paraphrases, negations, and ambiguous terms. Each query should have an expected evidence set, including cases where no answer is available.
The interesting result is not simply a higher average score. It is which failure classes improved and which became worse. A configuration that helps broad conceptual questions may bury a precise match under many similar passages.
Reranking can help sort candidates, but it adds another stage whose behavior and cost need measurement. Candidate selection and final ordering should be inspected separately.
Respect filters throughout
Tenant and permission filters belong in every retrieval path. A lexical branch and a vector branch that apply different access rules create a serious inconsistency, even if their final ranking looks excellent.
I also want the displayed citations to preserve document identity through fusion and reranking. The answer should not lose the connection to the source while the search pipeline rearranges candidates.
Hybrid search is valuable because it accepts that relevance has several signals. The engineering challenge is to combine those signals without hiding their tradeoffs. Better retrieval starts with understanding why each candidate was allowed into the conversation.