Qdrant Search Strategies

How to Improve Search Results with Advanced Strategies

These strategies complement basic vector search. Use them after confirming the embedding model is fitting the task and HNSW config is correct. If exact search returns bad results, verify the selection of the embedding model (retriever) first. If the user wants to use a weaker embedding model because it is small, fast, and cheap, use reranking or relevance feedback to improve search quality.

Missing Obvious Keyword Matches

Use when: pure vector search misses results that contain obvious keyword matches. Domain terminology not in embedding training data, exact keyword matching critical (brand names, SKUs), acronyms common. Skip when: pure semantic queries, all data in training set, latency budget very tight.

• Dense + sparse with prefetch and fusion Hybrid search
• Prefer learned sparse (miniCOIL, SPLADE, GTE) over raw BM25 if applicable (when user needs smart keywords matching and learned sparse models know the vocabulary of the domain)
• For non-English languages, configure sparse BM25 parameters accordingly
• RRF: good default, supports weighted (v1.17+) RRF
• DBSF with asymmetric limits (sparse_limit=250, dense_limit=100) can outperform RRF for technical docs DBSF
• Fusion can also be done through reranking

Right Documents Found But Wrong Order

Use when: good recall but poor precision (right docs in top-100, not top-10).

• Cross-encoder rerankers via FastEmbed Rerankers
• See how to use Multistage queries in Qdrant
• ColBERT and ColPali/ColQwen reranking is especially precise due to late interaction mechanisms, but it is heavy. It is important to configure and store multivectors without building HNSW for them to save resources. See Multivector representation

Right Documents Not Found But They Are There

Use when: basic retrieval is in place but the retriever misses relevant items you know exist in the dataset. Works on any embeddable data (text, images, etc.).

Relevance Feedback (RF) Query uses a feedback model's scores on retrieved results to steer the retriever through the full vector space on subsequent iterations, like reranking the entire collection through the retriever. Complementary to reranking: a reranker sees a limited subset, RF leverages feedback signals collection-wide. Even 3–5 feedback scores are enough. Can run multiple iterations.

A feedback model is anything producing a relevance score per document: a bi-encoder, cross-encoder, late-interaction model, LLM-as-judge. Fuzzy relevance scores work, not just binary (good/bad, relevant/irrelevant), due to the fact that feedback is expressed as a graded relevance score (higher = more relevant).

Skip when: if the retriever already has strong recall, or if retriever and feedback model strongly agree on relevance.

• RF Query is currently based on a 3-parameter naive formula with no universal defaults, so it must be tuned per dataset, retriever, and feedback model
• Use qdrant-relevance-feedback to tune parameters, evaluate impact with Evaluator, and check retriever-feedback agreement. See README for setup instructions. No GPUs are needed, and the framework also provides predefined retriever and feedback model options.
• Check the configuration of the Relevance Feedback Query API
• Use this as a helper end-to-end text retrieval example with parameter tuning and evals to understand how to use the API and run the qdrant-relevance-feedback framework: RF tutorial