qdrant-search-quality-diagnosis

Run each query twice, once with approximate HNSW search and once with exact=true, then compute recall@k from the overlap between result sets. You can also use the Qdrant Web UI ANN Recall tab for a no-code comparison.

Approximate search underperforms when HNSW parameters are too low. Increase hnsw_ef at query time, raise ef_construct to 200 or more, and set m to 32 for high recall workloads.

Yes, binary quantization causes severe quality loss without rescoring. Always enable rescore and use oversampling of 3-5x minimum to recover recall, and test the impact on your data before production.

Test the top MTEB models on 100-1000 sample queries from your domain and score them against a labeled set. Most search quality issues trace back to the embedding model rather than index configuration.

Filtered queries degrade when payload indexes are missing, because HNSW cannot traverse filtered-out nodes efficiently. Create payload indexes on filtered fields and consider the ACORN algorithm in Qdrant v1.16+ for filtered search.

Build a labeled query set from human annotation, search logs, or LLM-synthetic data, then score retrieval with ranx. Use Recall@k for RAG, MRR for single-answer search, or NDCG@k for re-ranking, and gate CI on a metric threshold.