Filtered ANN search is approximate-nearest-neighbour search that must also satisfy metadata conditions — returning only vectors that are both similar to the query and match a structured filter such as a category, date range, or owner. It is one of the most common and most technically demanding query types in production vector systems.
The difficulty is that filtering and approximate search pull against each other. ANN indexes like HNSW achieve their speed by following a carefully built graph, but a filter removes many nodes from consideration, which can break the graph’s navigability and cause the search to miss genuine matches — the recall cliff. Handling this well requires the filter to be applied during the search, not merely before or after it.
Different systems solve this with techniques like in-graph filtering, adaptive traversal, and filter-aware indexing, all aimed at keeping recall high without sacrificing speed. The quality of a database’s filtered ANN search is often a key differentiator, since real applications rarely want pure similarity — they want the most similar results that also meet concrete business constraints.