Ask what is needed: the top ten

Search results are ranked by a learned relevance model. The model is pairwise: it can only say which of two results is more relevant, and every comparison is a billed model call. Scores are never equal.

The ticket says: "rank(results) returns all results sorted best first." Here is the only code that calls it:

def homepage_widget(results):
    return [r["id"] for r in rank(results)[:10]]

The data

results = [{"id": "doc000001", "score": <Score>}, ...]
# a["score"] < b["score"] is one model call. So is >, ==, and so on.
# -s gives a Score that compares the other way round (no call).