Qdrant Hybrid Search Combining

Combining Prefetch Results

The outer query fuses ranked candidate lists from all parallel prefetches into one ranked list of results. Fusion methods differ in whether they use rank, score or directly vector representations of candidates (their similarity to the outer query) and whether final score incorporates payload metadata. All methods support flat (one fusion step) and nested (multi-stage) prefetch structures.

Tune in this order, cheapest first: confirm fusion beats each prefetch alone on your labels, compare Qdrant's default RRF vs DBSF, if tuning RRF: settle k, then sweep weights at that k, and validate on held-out queries. A weight pair is only valid for the k it was tuned with Tune in this order

Scores Are Not Comparable Across Prefetches & You Want Some Easy Baseline

Use when: searches produce scores on different scales, like BM25 and cosine on dense embeddings.

RRF

• RRF (Reciprocal Rank Fusion): rank-based, ignores scores magnitude.

• Start with RRF at Qdrant's defaults, k=2 and equal weights.

• Tune k to control rank sensitivity in RRF fusion, choosing the value from your labels Use labels to choose a k range.

• Add per-prefetch weights when one search should dominate, using Weighted RRF. Weights should be customized per collection and retrievers' score distributions. Tune weights last.

DBSF

• DBSF (Distribution-Based Score Fusion): normalizes score distributions per prefetch before fusing them, for that, instead of min-max, uses mean +- 3 deviations on prefetched list of scores. Avoid relying on resulting absolute scores, as scores in DBSF are normalized per prefetch (aka per a retrieved list of search results), and might be uncomparable across queries.

• DBSF has no parameters to tune, so compare it before hand-tuning RRF k and weights. Confirm on your own labels Compare RRF and DBSF on your labels

Need Custom Fusion

Use when: recency, popularity or other payload values should affect the merged ranking alongside candidate scores or you need a custom fusion.

With formula query, access score of each prefetch and, if desired, payload field values.

If you want to implement custom fusion on score of each prefetch:

• Use decay or any other available expressions for normalizing score distributions before fusing them.

• Parameters of these expressions should be based on the collection & retriever score distributions (for example, adjusting these parameters on a subsample of real queries).

• Formula query is unable to provide ranks for custom fusions

When using FormulaQuery over multiple prefetches (e.g. per-representation weighting):

• $score[i] indexes prefetches in declaration order. Reordering the prefetch= list silently shifts which weight applies to which retriever.

• Provide defaults for every $score[i] so the formula still evaluates for candidates that surfaced from only a subset of prefetches.

• On Qdrant 1.19.1 or newer, max and min expressions take the best or worst of several $score[i] (for example, keep each point's best score across per-field BM25 prefetches), cap a boost (min(cap, boost)), or clamp a value with max(lo, min(hi, x)). A point absent from a prefetch has $score[i] of 0.0, which wins max over negative scores and min over positive ones, so set defaults (keys include the $, placed beside formula, never inside it) to a value outside the expected score range such as -1e30 for max or 1e30 for min Max and min.

• On 1.19.1 or newer, acosh supports exact geodesic-distance rescoring for hyperbolic embeddings (Lorentz model with Dot distance, acosh(-$score), negated so closest ranks first). The input must be at least 1 Hyperbolic distance.

• Start with RRF when scores are on incomparable scales (e.g. BM25 + cosine). Reach for FormulaQuery only when explicit per-representation weighting or payload-driven boosts are required, and normalize each $score[i] (decay or min-max on a sampled distribution) before combining linearly.

Need Good Ranking of Fused Candidates and Ready To Spend More Resources

Use when: you want to use similarity between query and candidates' vector representations as the prefetches combiner and simultaneously ranker. More resource heavy than score/rank based fusions, but might be necessary due to use case requirements or need in a high top-K precision of results (when parallel prefetches have overall a good recall of retrieved candidates).

More candidates only help if the ranker can use them. Deeper prefetches can raise the best possible score

You can use any type of vector as an outer query over the prefetches, to perform the fusion on the server-side in one QueryAPI request: sparse, dense, multivector. For that, same type of vector representations for documents need to be stored as named vectors per point.

Instead of using client-side fusion through cross-encoders, a popular option is Late interaction models-based fusion, through reranking on multivectors (e.g. ColBERT for text, ColPali and ColQwen for images).

• Most precise but highest compute/resource usage.

• Configure multivectors used for fusion through reranking with HNSW disabled like in Hybrid Search with Reranking tutorial.

What NOT to Do

• Use linear weighted fusion on incomparable score ranges. Why not.

• Use "vibe" defined weights in weighted RRF. Weights should be fine-tuned per dataset and retrieval pipelines.

• Pick any fusion type without comparative experiments.

• Use late interaction multivectors for fusion without evaluating cheaper analogues, for example, MUVERA. More in multi-vector Qdrant search course