RAG & Retrieval · Ground answers in evidence
How do you combine BM25 and vector search?
Short answer
Run lexical and vector retrieval, then fuse the ranked lists using reciprocal rank fusion or tuned weights. Apply the same filters to both and rerank the merged candidates.
Lexical search is precise for exact terms and identifiers; vector search handles paraphrases. Combining them covers more query types than either alone. The engineering challenge is merging two rankings that use different score scales.
- Reciprocal rank fusion is a robust default because it uses ranks, not raw scores.
- Tune weights on real queries if you need score-level fusion.
- Apply metadata and permission filters consistently to both branches.
- Rerank the fused set to produce the final ordering.
Follow-ups you should expect
Interviewers use these to check whether the first answer was understanding or memorisation.
- 1How would you evaluate whether hybrid search helped?
- 2Why is rank fusion more stable than score addition?
- 3What would you do if the lexical branch returns nothing?
Say it in the interview
I run both retrievers in parallel and fuse the results, usually with reciprocal rank fusion because it avoids comparing incompatible scores. The same permission filters apply to both. Then I rerank the merged candidates. I justify the extra complexity with real queries that vector search misses, like product codes or error strings, and I measure the lift rather than assuming hybrid is always better.
Use this as a model, not a script. Saying it in your own words is what makes it survive follow-up pressure.
Related questions
Reading is step one
Now say it out loud.
Understanding an answer and delivering it under pressure are different skills. Rehearse this topic in a live RAG & Retrieval mock interview and get scored feedback on depth, clarity and confidence.