qdrant-search-quality

Start by comparing exact search against approximate HNSW search to isolate the cause. If exact search is also bad, the embedding model or data chunking is at fault; if only approximate search is bad, tune hnsw_ef, ef_construct, and m parameters.

Use hybrid search with the Query API prefetch mechanism, running a sparse vector search (BM25 or SPLADE) alongside dense vector search in parallel. Fuse the results with RRF as a baseline, or DBSF when score distributions need normalization.

RRF is rank-based and ignores score magnitudes, making it a solid default when scores are on incomparable scales like BM25 and cosine. DBSF normalizes score distributions per prefetch using mean and standard deviation before fusing, but its absolute scores are not comparable across queries.

Yes, especially binary quantization without rescoring, which causes severe quality loss. Enable oversampling (3-5x minimum for binary) with rescoring, and always measure recall impact on your own data before deploying quantization to production.

Build a labeled golden set of queries with known relevant documents, then score retrieval with metrics like recall@k for RAG, MRR for single-answer search, or NDCG for ranking. Use the ranx library and gate CI on metric thresholds to catch regressions.

Use Qdrant's Relevance Feedback API when the dense retriever misses relevant items outside the initial ANN pool or when reranking a large candidate set is too costly. It requires calibrating formula weights on 50-200 representative queries with a consistent feedback model.