"""
DSPy RAG Evaluation Module

This module provides tools for evaluating the Heritage RAG system:
- Golden dataset loading
- Metric functions (count accuracy, slot extraction)
- Evaluation harness for running tests against the API
"""

from .metrics import (
    count_accuracy,
    slot_extraction_accuracy,
    heritage_rag_metric,
)
from .dataset_loader import load_golden_dataset, GoldenExample
from .run_evaluation import run_evaluation

__all__ = [
    "count_accuracy",
    "slot_extraction_accuracy", 
    "heritage_rag_metric",
    "load_golden_dataset",
    "GoldenExample",
    "run_evaluation",
]