| from agentic_search.evaluation.aggregate import latest_records | |
| def test_latest_records_replaces_a_failed_attempt_with_a_retry() -> None: | |
| records = [ | |
| {"example_id": "q1", "system": "sac", "status": "failed"}, | |
| {"example_id": "q1", "system": "sac", "status": "completed"}, | |
| {"example_id": "q2", "system": "fixed", "status": "completed"}, | |
| ] | |
| latest = latest_records(records) | |
| assert len(latest) == 2 | |
| assert next(row for row in latest if row["example_id"] == "q1")["status"] == "completed" | |