curious what the angle is here. most similarity/search tools optimize for pairwise comparisons. what advantage do you get by moving to group-level reasoning