What is pre-filtering versus post-filtering in vector search?

Pre-filtering applies metadata filters before vector search, restricting the candidate set the ANN index must consider. Post-filtering runs vector search first, then filters the top-k results by metadata, which is simpler but can return too few results when the filter is highly selective. Pre-filtering is efficient when metadata fields are indexed but requires the ANN algorithm to handle filtered traversal, which can degrade recall if implemented naively. Qdrant and Weaviate use tight integration of pre-filtering with HNSW to maintain recall under filters.