Function main

python/agents/invoice-processing/eval/eval.py:1418–1459

()

def main():
    args = _parse_args()

    master = _load_master(args.master_data)

    gt_dir, agent_dir = _resolve_directories(args)

    case_ids = _collect_case_ids(args, gt_dir, agent_dir)

    # Initialize LLM if needed
    llm_model = None
    if not args.skip_llm:
        llm_model = _init_llm()
        if llm_model is None:
            print(
                "\n  Warning: LLM unavailable (missing packages or PROJECT_ID). "
                "Running deterministic only."
            )
            print(
                "  Install: pip install google-cloud-aiplatform python-dotenv"
            )

    domain_name = master.display_name if master else "Invoice Processing"
    print("=" * 70)
    print(f"{domain_name.upper()} AGENT EVALUATION")
    print("=" * 70)
    print(f"  Ground truth: {gt_dir}")
    print(f"  Agent output: {agent_dir}")
    if master:
        print(f"  Master data:  {master.source_path}")
    print(f"  Cases:        {len(case_ids)}")
    print(f"  Tolerance:    ${args.tolerance}")
    print(f"  LLM judge:    {'enabled' if llm_model else 'disabled'}")

    results = _run_evaluation_loop(
        case_ids, gt_dir, agent_dir, args.tolerance, llm_model, master
    )

    stats = compute_aggregate_stats(results, master)
    print_report(stats, results)

    _save_results(args, stats, results, gt_dir, agent_dir, llm_model, master)


if __name__ == "__main__":