cli.py

import os
import sys

from core.dataset_loader import load_dataset
from core.agent_loader import load_agent
from evaluator import evaluate

import json
from pathlib import Path
from datetime import datetime

def main():

    if len(sys.argv) != 3:
        print("Usage:")
        print("python cli.py <agent_file.py> <dataset.json>")
        sys.exit(1)

    agent_file = sys.argv[1]
    dataset_file = sys.argv[2]

    document_text, documentId, samples = load_dataset(dataset_file)

    run_agent = load_agent(agent_file)

    scores = evaluate(documentId, document_text, samples, run_agent)

    print("\n===== Evaluation Results =====\n")
    current_datetime = datetime.now().strftime("%Y-%m-%d_%H-%M-%S")
    results_path = Path("results", f"results-{Path(agent_file).stem}-{Path(dataset_file).stem}-{current_datetime}.json")

    # Ensure the results directory exists
    os.makedirs(results_path.parent, exist_ok=True)

    with open(results_path, "w") as f:
        json.dump(scores, f, indent=2)

    print(f"\nResults saved to {results_path}")

if __name__ == "__main__":
    main()