qdrant-hybrid-search-combining

Use the outer query to fuse ranked candidate lists from parallel prefetches. Start with RRF when scores are on incomparable scales like BM25 and cosine, or use DBSF to normalize score distributions per prefetch before fusing.

RRF is rank-based, ignores score magnitudes, and works as a decent default when scores are incomparable. DBSF normalizes each prefetch's score distribution using mean plus or minus three deviations, but its absolute scores are not comparable across queries.

Use Weighted RRF to assign per-prefetch weights when one search should dominate. Weights must be tuned per collection and retriever score distributions rather than chosen by intuition.

Yes, use FormulaQuery to access each prefetch's score and payload values in a custom expression. Provide defaults for every score index and normalize each score with decay or min-max before combining linearly.

Use late-interaction multivector reranking (ColBERT, ColPali, ColQwen) when you need high top-K precision and prefetches already have good recall. It is the most precise but most resource-heavy option, so evaluate cheaper alternatives like MUVERA first.

Linear weighted fusion fails when prefetch scores are on incomparable ranges, such as BM25 versus cosine similarity. Normalize each score distribution first or switch to rank-based RRF before applying any weights.