qdrant-vector-search

Use a must/should/must_not filter combined with the vector query, storing the filtered fields in the point payload. Call create_payload_index on frequently filtered fields, since unindexed payload filters degrade to a linear scan over candidates.

Raise the query-time ef parameter (hnsw_ef in some clients) for queries where recall matters more than latency, without rebuilding the index. The m and ef_construct parameters trade build time and memory against recall at index construction.

Qdrant ships built-in scalar, binary, and product quantization. Scalar int8 quantization is the safe default recall/memory tradeoff; binary quantization is more aggressive and suits embedding models designed to tolerate it. Re-measure recall@K after enabling it.

Use named vectors to store multiple vector fields per point, such as a dense embedding plus a sparse keyword vector, in one collection. Qdrant's native multi-vector query then combines both scores server-side.

A cache in front of the fusion pipeline without an invalidation strategy serves stale fused results indefinitely. Decide up front whether the cache is TTL-based, invalidated on write, or eventually consistent, and test that decision explicitly.

Each backend needs its own SQL dialect branch, and a test suite exercising only SQLite can pass while the Postgres branch drifts. Force the other backend in tests via a real connection object rather than overriding a read-only backend property.