KNN prefiltering applies attribute filters during vector search, not after it. This helps return the nearest matching documents within the filtered subset, which is especially important for tenant-aware, category-aware, and permission-filtered search.