Qdrant Hybrid Search Prefetches

Different Searches in One Query API Request

Each prefetch runs exactly one search per one query.

Understand if user wants to run several parallel searches on:

• The same vector representations but different queries or filters.
• Different vector representations but the same raw query.

If first, help user to design logic of constructing query or/and filters on application side and then check Combining Searches. Don't forget to create indices on filterable payload fields, immediately after collection creation, prior to building HNSW, so filterable HNSW could be constructed.

If second, use named vectors, which allow to store multiple vector types per point in one collection. Beware that named vectors currently can be configured only at collection creation. To choose vectors, check following recommendations.

Missed Keyword Matches

Use when: pure vector search misses exact term or keyword matches and you need lexical retrieval alongside semantic search.

Most likely you need a sparse vector for exact text search alongside the dense one. Qdrant uses sparse vectors for lexical searches, as payload filtering doesn't provide any ranking score.

Choose a Sparse Vector for Text

• BM25 statistical representations, built into Qdrant core (computed server-side). Good baseline, works out-of-domain, usually for long texts. Can be used for non-English content, but needs to be configured per language (tokenization, stemming, stopwords, etc) at indexing and retrieval time. More in Text Search Guide
• BM42 learned sparse, based on BM25, but better for small chunks of text & with meaning understanding. Works only on English. Requires fine-tuning for domain-specific retrieval. Requires FastEmbed (Python/REST only, not available in all SDKs). Not maintained.
• miniCOIL learned sparse, BM25 with additional understanding of words meaning in context. Works only on English. Requires fine-tuning for domain-specific retrieval. Requires FastEmbed. Usage shown in FastEmbed miniCOIL documentation.
• SPLADE++ learned sparse with term expansion. Heavier inference and resources usage but better performance due to term expansion. Requires fine-tuning for domain-specific retrieval. Provided in Qdrant Cloud Inference and FastEmbed versions work only on English. To use with FastEmbed, check FastEmbed SPLADE documentation.
• External learned sparse embeddings, for example BAAI/bge-m3.

What to remember when using sparse vectors for lexical search:

• tokenization and stemming affect exact matches, especially on custom codes, terms, etc.

What to remember when using Qdrant BM25 and miniCOIL (based on BM25):

• avg_len in formula is not computed server-side, it is a user responsibility and passed as a parameter. Calibrate per field — defaults assume document-length text; short fields (titles, tags) need a much smaller value or BM25 scoring is skewed (avg_len=256 against a 10-word title overweights term frequency).
• BM25 might be not good for small chunks of text, as BM25 algorithm was initially created for search on long documents; consider adjusting document statistics in sparse vectors (TF & IDF, k, b).
• Qdrant BM25 vectors are configured per language, so consider customizing stop words, stemming & tokenization when users documents mix several languages or carefully configure vectors per point when they are monolingual.

More on Sparse Vectors for Text Search

Need to Combine Multiple Representations of the Same Item

Use when: the same item is embedded in multiple ways (e.g. different models, languages, modalities, or different fields like title/abstract/chunk) and you want to search across different representations in one request (don't have to be all of them, can be even one).

Use multiple named vector prefetches, each prefetch covers one representation.

A representation only earns its own prefetch if it carries signal independent of the others — e.g. title vocabulary the body never repeats, or an abstract treated as a single semantic unit vs. individual chunks. Don't add a prefetch per field reflexively; verify each candidate contributes content the other vectors don't.

When a representation's signal is mostly lexical — keyword-driven titles, codes, tags, or other short fields — prefer a sparse named vector (e.g. BM25) over an additional dense embedding. Server-side BM25 in Qdrant avoids the inference cost of another dense model and stores far less per point. Skip this when the field carries paraphrase or conceptual signal that exact-term matching would miss.

• End-to-end worked example fusing title, abstract, chunk, and sparse-title named vectors with RRF and document-level grouping in one Query API call: Multi-Representation Search tutorial
• If you have groups and subgroups of representations (document -> chunk, image -> patch), you could use searching in groups. To not store identical payloads several times, check Lookup in Groups. Index the grouping payload field (e.g. document_id) as a keyword payload index before grouping.
• When grouping chunk-level points back to documents, each prefetch only contributes the candidates it returned — so size per-prefetch limit well above the final document limit (rule of thumb: prefetch_limit ≥ final_limit × expected_chunks_per_document), otherwise a few documents with many chunks saturate the candidate pool and relevant documents drop silently. Validate grouped recall on a labeled sample.
• When per-document vectors (title, abstract) would be duplicated across every chunk-level point, the duplication can dominate storage at scale. Keeping them denormalized in one collection makes queries simpler (single Query API call, every representation reachable from any point); a sidecar collection joined via Lookup in Groups is the alternative when storage matters.

You can also search directly on multivectors, a matrix of dense vectors, in a prefetch.

However, it comes with several considerations, as multivectors were designed to support late interaction models using max similarity metric, so it's impossible to retrieve the list of individual max similarity scores for each query vector.

Moreover, multivectors are rarely a good pick for prefetch:

• max similarity metric is not symmetric, so using HNSW index with it could be problematic
• multivector representations are very heavy, as search process on them.

There are ways to make multivector retrieval cheaper (MUVERA, pooling), you can see more in "Evaluating Tradeoffs of Multi-stage Multi-vector Search"

What NOT to Do

• Choose any search method (for example, BM25) without evaluation of its quality & resources used.
• Use any search method (for example, BM25) without paying attention to the specifics of their configuration and applicability to the use case.