`aip_sdk.RowCoverage`

aip_sdk.RowCoverage(total_rows: int, answered_rows: int, unanswered_rows: int, coverage: float, unanswered_breakdown: list[UnansweredRows] = list[UnansweredRows](), scores_exclude_unanswered: bool = True)

How much of the dataset the system under test actually answered.

Every score in metrics is computed over answered_rows, not total_rows: a row the endpoint refused carries no response to evaluate, so it is left out of the metric rather than scored as a failure. Read coverage alongside any score — a run at 0.38 coverage reports a number about 38% of the dataset.

Attributes

total_rows int: Rows in the dataset the run was given.
answered_rows int: Rows the endpoint answered, and the denominator of every metric.
unanswered_rows int: Rows the endpoint never answered, excluded from every metric.
coverage float: answered_rows / total_rows, between 0.0 and 1.0.
unanswered_breakdown list[UnansweredRows]: Why the unanswered rows went unanswered, largest group first. Empty when every row was answered.
scores_exclude_unanswered bool: Whether this run's scores really do leave the unanswered rows out. False for a run scored before the platform excluded them, whose metrics still count every refused row as a failure — read its scores as covering total_rows, and re-run it for figures that cover answered_rows.