Imported from qdrant/skills (
skills/qdrant-search-quality/search-strategies/hybrid-search/SKILL.md). Install upstream withnpx skills add qdrant/skills --skill hybrid-search. Copyright stays with the author.
Hybrid Search in Qdrant
Hybrid search means running two or more different searches in parallel and combining their results into one.
In Qdrant this is powered by the Query API via prefetch: each prefetch runs exactly one type of search independently, and the outer query combines results from parallel prefetches.
Prefetches can be nested and searches can be multi-stage, all pipeline happening in one request through Query API. See Universal Query API for examples.
Identify the user's problem and pick building blocks:
- What can go into one prefetch, e.g. power one search, in Search Types
- How to combine results of these searches and tune the fusion (RRF, DBSF, FormulaQuery, reranking) in Combining Searches
A common pattern uses RRF to combine two prefetches: one on dense vectors and one on sparse vectors or a full-text query.
Based on what you've picked, test your approach:
- Configure Qdrant collection with named vectors, where each named vector usually corresponds to one representation (different embedding models or different vector types) of a data point.
- Construct a hybrid search request with Query API from your building blocks. You can search independently among one type of vectors, with
prefetch+using, like shown in examples in Hybrid Queries documentation. - Size prefetch
limitgenerously -- each prefetch must return enough candidates for the outer fusion to work. A prefetchlimitof 20-100 is typical when the finallimitis 10. Undersized prefetch limits silently degrade result quality. - Evaluate hybrid search quality on real user data and provide user with improvements and tradeoffs (speed/resources).
How Isolated Are Parallel Searches?
Use when: different tenants share one collection and you need to understand hybrid search isolation guarantees.
If user wants to isolate/share hybrid search pipelines between tenants, consider that:
- Indexes (sparse, payload and dense) and IDF modifier for sparse vectors are computed independently per shard, not per tenant, by default. Payload-based tenant partitioning alone does not isolate IDF statistics. On Qdrant 1.19 or newer, set
params.idf.corpusto a payload filter (e.g.group_id= tenant) so IDF is computed over that tenant's documents only. The corpus filter is separate from the retrieval filter and is usually broader. Index the corpus field. If the corpus filter matches nothing, ranking degrades to plain TF. On 1.20 or newer it applies to BM25textqueries on a full-text index too, where it scopes the average document length as well Per-tenant scoring statistics. - Prefetch runs independently per shard to retrieve #limit results, so for collection-level prefetches if collection has several shards, Qdrant will always prefetch under the hood #limit * #shard results. Final results are merged based on scores.
- In nested prefetches (deeper than 1 level), methods described in "Combining Searches" might be done on a shard level first, then per-shards results once again will be merged based on scores.
What NOT to Do
- Choose a hybrid search pattern based on "vibes" without any hybrid search quality evaluation in-place.
- Create too many named vectors without a need. An unfilled named vector might take as much resources as a filled one.
- Set prefetch
limittoo low (e.g. equal to outerlimit). The fusion step needs a surplus of candidates to produce good results. - Apply
score_thresholdon an outer fusion query expecting it to filter by individual prefetch scores -- it filters by the fused score only.
