qdrant-search-strategies

First diagnose the symptom: missing keyword matches need hybrid search with sparse vectors, good recall but poor precision needs multistage reranking, and redundant results need MMR. Each symptom requires its own strategy rather than one blanket fix.

Use hybrid search via the Query API with prefetch: run a sparse vector search (BM25, SPLADE, or miniCOIL) and a dense vector search in parallel prefetches, then fuse results with RRF or DBSF. Store each representation as a named vector in the collection.

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

Relevance feedback modifies the vector search itself using scores from a feedback model on a few seed documents, surfacing relevant items outside the initial ANN pool. Use it when reranking is too costly or the dense retriever misses relevant items, after calibrating formula weights on 50-200 real queries.

By default, IDF statistics are computed per shard, not per tenant, so tenant vocabularies get merged and distort scoring. On Qdrant 1.19 or newer, the idf search parameter can scope statistics to a payload-filtered corpus for per-tenant isolation.

Do not add hybrid search before verifying pure vector search quality, since it adds complexity and can mask embedding model problems. Also avoid choosing fusion methods or weights without evaluating them on real query data from your collection.