Explain the execution flow of a filtered vector search query in Qdrant?

A filtered vector search combines HNSW graph traversal with payload condition checking in a single, integrated pass, rather than running two entirely separate operations and merging their results afterward.

Qdrant's query planner looks at the filter's expected selectivity and whether a payload index exists for the filtered fields, choosing between letting the payload index narrow the candidate set upfront (efficient for highly selective filters) versus checking the filter condition as the HNSW graph is traversed (efficient when the filter is broad and most points would match anyway).

This adaptive strategy is what avoids the two failure modes of a naive approach: pure post-filtering (searching first, then filtering, risking too few final results) and pure pre-filtering via a full unindexed scan (slow if the payload isn't indexed) — Qdrant instead picks whichever strategy fits the specific combination of filter selectivity and available indexes for that query.