Adaptive filtered traversal is a query-execution strategy for graph-based vector indexes that changes how it navigates the index depending on how restrictive the active metadata filter is. Rather than applying one fixed approach to every query, it measures the selectivity of the filter — what fraction of vectors pass — and adapts accordingly.
When a filter is loose and most vectors qualify, the engine behaves like ordinary approximate search, walking the graph normally. When a filter is highly selective and only a tiny fraction of vectors qualify, a normal graph walk would waste most of its steps on excluded nodes. In that case the engine may switch to a brute-force scan of the small qualifying set, or bias its traversal heavily toward nodes likely to pass the filter.
This adaptivity is what prevents the recall cliff: the sudden collapse in result quality or spike in latency that occurs when a filter is too tight for a static algorithm to handle. By matching the execution strategy to the filter, adaptive traversal keeps both speed and recall stable across the full range of query selectivity.