How would you build a hybrid search query that retrieves candidates from both a dense embedding and a BM25-style sparse vector, then fuses the results?
Use two prefetch branches and fuse their candidate lists with RRF
I would model the collection with a named dense vector and a named sparse vector, generate both query representations, retrieve a candidate set from each using prefetch, and then apply a fusion query such as Reciprocal Rank Fusion. RRF is a good starting point because dense and sparse scores do not necessarily share the same numerical scale. The important design decision is candidate depth: each branch should retrieve enough candidates that relevant documents are unlikely to disappear before fusion, while not making the first-stage search unnecessarily expensive. I would then measure recall, latency, and fusion quality offline before tuning limits. A common mistake is multiplying or directly adding raw dense and sparse scores without calibration. The exact Universal Query API and client syntax is version-sensitive, so production code should match the installed Qdrant SDK/server release.
- 1
RRF uses ranking positions rather than requiring dense and sparse scores to be directly comparable.
- 2
Candidate depth is a quality/latency knob: too few candidates can reduce recall before fusion; too many increase work.
- 3
An alternative is DBSF or a calibrated weighted score when you have evidence that score distributions are stable and comparable enough for your workload.
- 4
Qdrant's Universal Query API, prefetch, and fusion interfaces are version-sensitive; verify the exact SDK syntax for the deployed client/server version.
Scenario Questions
- 1
You have dense and sparse candidate lists. Why should you fuse the rankings instead of returning whichever list is longer?
- 2
Your hybrid query returns only dense-like results. What configuration would you inspect first?
- 1
A hybrid search returns good results but misses rare exact matches. What candidate-generation or fusion parameters would you investigate?
- 2
Why can increasing both prefetch limits improve recall while also increasing latency?
- 1
Design a Qdrant hybrid query for 50 million documents with dense semantic retrieval and BM25-style sparse retrieval. How would you tune candidate depth?
- 2
Your RRF result quality is good but a reranker receives too many candidates. How would you balance first-stage recall against reranking cost?
- 1
Design an adaptive hybrid retrieval strategy where candidate depth changes by query type or confidence while meeting a strict p95 latency target.
- 2
How would you compare RRF, DBSF, and learned fusion using offline relevance judgments and online A/B testing?
Follow-up Questions
- ❓How would you choose the dense and sparse candidate limits before RRF fusion?
- ❓When would DBSF or a learned fusion method be preferable to RRF?