Filtered Approximate Nearest Neighbor Search in Vector Databases: System Design and Performance Analysis

A vector query often has two requirements: find semantically nearby records and return only records satisfying a metadata condition. The nearest records in the entire corpus may be invalid for that condition. Searching first and filtering afterward can therefore return too few results or miss the nearest valid ones.

This paper studies filtered approximate nearest-neighbor search inside FAISS, Milvus and pgvector. It separates filtering strategies, introduces a movie-and-review dataset with structured metadata, and examines how index parameters and system execution policies affect the recall-latency tradeoff.

Its Global-Local Selectivity metric compares the fraction of eligible records near a query with their fraction across the corpus. Two filters can accept the same number of records while creating very different search difficulty if one has eligible points nearby and the other does not.

The February 2026 experiments show why an index-family label is insufficient. Query planning, adaptive search and exact fallbacks can matter as much as the underlying index. The authors' benchmark extension provides the starting point for reproduction. The findings concern specific versions and an in-memory scalar-filter workload, not every current configuration of the named systems.

A customer may report that search works in a demo but fails after adding account, date or product filters. Before changing the embedding model, inspect whether the vector-search plan finds enough eligible candidates. The relevant document may exist and have a good vector, yet never survive the chosen execution path.

This is also a measurement problem. Recall must be computed against the nearest valid records, not the unfiltered top results. Otherwise a benchmark can reward a system for finding records the application is not allowed or intended to return.

Use vector databases and RAG evaluation together. Establish a filtered exact-search reference, then compare approximate plans at a required quality level. The goal is reliable evidence under real constraints, rather than the highest query rate for a different, unfiltered problem.