An alternative, but related, approach that would eliminate false positives could be to do an N-ary tree instead, where the bloom filter in this solution is just the first level of the tree. You could then download a json document for the child of the matching top level element to narrow it down to specific documents. You could also calculate relevancies this way.