I wonder how more accurate this is compared to just doing embedding similarity of the query vector and the category labels