Filtered vector search is similarity search constrained by metadata conditions, so that results must satisfy both a semantic-similarity requirement and structured filters such as category, price range, date, language, or user ID. It reflects how vector search is actually used in practice, where pure similarity is rarely enough on its own.
For example, an e-commerce search might ask for products visually similar to an uploaded image, but only those in stock, under a certain price, and in a chosen category. The vector component handles the visual similarity; the filters enforce the hard business rules. Combining the two in one query is what makes vector search useful for real applications.
How a database executes the filter — before the search (pre-filtering), after it (post-filtering), or during the graph traversal (in-graph filtering) — has a large effect on both speed and recall. Strong filtered vector search is a hallmark of mature vector databases, because doing it naively leads to either slow queries or missing results when filters are restrictive.