With the knn predicate, the score and similarity (distance) can be derived from one another. It would be easier for a user to just use a score filter instead of trying to convert that value through a score-distance formula.

OpenSearch has two dedicated predicate options for this : max_distance and min_score as for the Elasticsearch or the Lucene backends: we can convert the values behind the scenes before passing it to the query.

See also https://opensearch.org/blog/vector-radial-search/