We should support threshold-based similarity search instead of solely topK search. This would allow us to more efficiently gather results by considering a relevancy threshold within the prefilter.