Build the exact iterative-filter query: pre-filter on a B-tree index, then compute
VECTOR_DISTANCE per surviving row in a CTE. Bypasses DiskANN entirely — used when a filter's
matching set is small enough that brute force beats ANN's non-convergence on disjoint clusters.
The vectorColumn IS NOT NULL guard keeps null-embedding rows out. @p0 is the query vector.
WITH (NOLOCK) is an intentional perf choice for a read-only similarity scan — it permits dirty
reads, acceptable for ranked search where exact transactional consistency isn't required.
Build the exact iterative-filter query: pre-filter on a B-tree index, then compute
VECTOR_DISTANCEper surviving row in a CTE. Bypasses DiskANN entirely — used when a filter's matching set is small enough that brute force beats ANN's non-convergence on disjoint clusters. ThevectorColumn IS NOT NULLguard keeps null-embedding rows out.@p0is the query vector.WITH (NOLOCK)is an intentional perf choice for a read-only similarity scan — it permits dirty reads, acceptable for ranked search where exact transactional consistency isn't required.