Function compute_aggregate_stats

python/agents/invoice-processing/eval/eval.py:1024–1067

Compute aggregate statistics across all evaluated cases. When master data is provided, dynamically computes per-group and per-field accuracy from eval_schema.comparison_groups instead of hardcoded field names.

(
    results: list[dict], master: MasterData | None = None
)

def compute_aggregate_stats(
    results: list[dict], master: MasterData | None = None
) -> dict:
    """Compute aggregate statistics across all evaluated cases.

    When master data is provided, dynamically computes per-group and per-field
    accuracy from eval_schema.comparison_groups instead of hardcoded field names.
    """
    evaluated = [r for r in results if r.get("status") != "ERROR"]
    errors = [r for r in results if r.get("status") == "ERROR"]

    if not evaluated:
        return {"total": len(results), "evaluated": 0, "errors": len(errors)}

    total = len(evaluated)
    passed = sum(1 for r in evaluated if r["all_correct"])

    decision_correct = sum(1 for r in evaluated if r["decision"]["match"])
    decision_matrix = _compute_decision_matrix(evaluated)

    if master:
        group_accuracy, field_stats = _compute_group_and_field_stats_schema(
            evaluated, master, total
        )
    else:
        group_accuracy, field_stats = _compute_group_and_field_stats_legacy(
            evaluated, total
        )

    llm_stats = _compute_llm_stats(evaluated)

    return {
        "total": len(results),
        "evaluated": total,
        "errors": len(errors),
        "passed": passed,
        "failed": total - passed,
        "pass_rate": passed / total,
        "decision_accuracy": decision_correct / total,
        "decision_matrix": decision_matrix,
        "group_accuracy": group_accuracy,
        "field_stats": field_stats,
        "llm_alignment": llm_stats,
    }


def _print_decision_stats(stats: dict) -> None: