Rerank

Agent calls POST /v2/rerank with a query and a list of candidate documents but omits top_n, then truncates the returned list client-side to 3.

Send query, documents, and top_n=3 so the API returns only the top-ranked subset (each result carries an index into the input documents and a relevance_score). Rely on the returned index→document mapping rather than re-slicing the original list by position.

A /v2/rerank response returns results with index and relevance_score. The agent assumes the results are already aligned to its original document array positions.

Map each result back to the source document using the result's index field, since results are returned in ranked (not input) order. Reorder/select source documents by index; never assume positional alignment.

Operator applies a fixed threshold of 0.5 on relevance_score and treats it as a calibrated probability of relevance across all queries.

Treat relevance_score as a within-query ranking signal, not a cross-query calibrated probability. Use relative ordering / top_n selection rather than a single global cutoff; if a cutoff is needed, tune it per use case and validate empirically. Any '0.5 means relevant' claim is [REQUIRES-VERIFICATIO…