triage/analysis/evaluate_metrics_analysis.py

#!/usr/bin/env python3
"""
Evaluate智能体评估指标分析脚本
用于统计evaluate的所有维度分数并绘制折线图
"""

import json
import os
import numpy as np
import matplotlib.pyplot as plt
from pathlib import Path
from typing import Dict, List
from file_filter_utils import filter_complete_files, print_filter_summary


def load_workflow_data(data_dir: str, output_dir: str = "", limit: int = 5000) -> List[Dict]:
    """
    加载工作流数据
    
    Args:
        data_dir: 数据目录路径
        output_dir: 输出目录路径（用于文件过滤）
        limit: 限制加载的病例数量
    
    Returns:
        工作流数据列表
    """
    workflow_data = []
    
    # 获取所有jsonl文件
    all_files = sorted(Path(data_dir).glob("*.jsonl"))
    
    # 过滤出完成的文件
    if output_dir:
        all_files = [str(f) for f in all_files]
        filtered_files = filter_complete_files(all_files, output_dir)
        filtered_files = [Path(f) for f in filtered_files]
        print_filter_summary(output_dir)
    else:
        filtered_files = all_files
    
    # 限制文件数量
    jsonl_files = filtered_files[:limit]
    
    print(f"将处理 {len(jsonl_files)} 个完成的文件")
    
    for file_path in jsonl_files:
        try:
            with open(file_path, 'r', encoding='utf-8') as f:
                workflow = []
                for line in f:
                    try:
                        data = json.loads(line.strip())
                        workflow.append(data)
                    except json.JSONDecodeError:
                        continue
                
                if workflow:
                    workflow_data.append(workflow)
        except Exception as e:
            print(f"加载文件 {file_path} 时出错: {e}")
    
    return workflow_data


def extract_evaluate_scores(workflow: List[Dict]) -> List[Dict]:
    """
    提取evaluate评分数据
    
    Args:
        workflow: 单个工作流数据
    
    Returns:
        evaluate评分列表
    """
    evaluate_scores = []
    
    for step in workflow:
        if step.get('agent_name') == 'evaluator' and 'output_data' in step:
            output_data = step['output_data']
            # 检查是否包含评估分数
            if any(key in output_data for key in [
                'clinical_inquiry', 'communication_quality',
                'information_completeness', 'overall_professionalism',
                'present_illness_similarity', 'past_history_similarity',
                'chief_complaint_similarity'
            ]):
                evaluate_scores.append(output_data)
    
    return evaluate_scores


def calculate_metrics_by_step(workflow_data: List[List[Dict]]) -> Dict[str, List[float]]:
    """
    计算每一步的评估指标平均值
    
    Args:
        workflow_data: 所有工作流数据
    
    Returns:
        各维度指标按步骤分组的平均值
    """
    # 找出最大步骤数
    max_steps = 0
    for workflow in workflow_data:
        evaluate_scores = extract_evaluate_scores(workflow)
        max_steps = max(max_steps, len(evaluate_scores))
    
    # 初始化数据收集器
    metrics_data = {
        'clinical_inquiry': [[] for _ in range(max_steps)],
        'communication_quality': [[] for _ in range(max_steps)],
        'information_completeness': [[] for _ in range(max_steps)],
        'overall_professionalism': [[] for _ in range(max_steps)],
        'present_illness_similarity': [[] for _ in range(max_steps)],
        'past_history_similarity': [[] for _ in range(max_steps)],
        'chief_complaint_similarity': [[] for _ in range(max_steps)]
    }
    
    # 收集每个步骤的评分
    for workflow in workflow_data:
        evaluate_scores = extract_evaluate_scores(workflow)
        
        for step_idx, score_data in enumerate(evaluate_scores):
            # 提取各维度分数
            for metric in metrics_data.keys():
                # 向后兼容性处理：将旧的 multi_round_consistency 映射到新的 information_completeness
                actual_metric = metric
                if metric == 'information_completeness' and metric not in score_data and 'multi_round_consistency' in score_data:
                    actual_metric = 'multi_round_consistency'
                
                if actual_metric in score_data and isinstance(score_data[actual_metric], dict):
                    score = score_data[actual_metric].get('score', 0.0)
                    metrics_data[metric][step_idx].append(score)
    
    # 计算平均值
    result = {}
    for metric, step_data in metrics_data.items():
        result[metric] = []
        for scores in step_data:
            if scores:
                result[metric].append(np.mean(scores))
            else:
                result[metric].append(0.0)
    
    return result


def plot_metrics_curves(metrics_data: Dict[str, List[float]], output_dir: str):
    """
    绘制评估指标折线图
    
    Args:
        metrics_data: 各维度指标数据
        output_dir: 输出目录
    """
    plt.figure(figsize=(16, 10))
    
    steps = list(range(1, len(next(iter(metrics_data.values()))) + 1))
    colors = ['#FF6B6B', '#4ECDC4', '#45B7D1', '#96CEB4', '#FECA57', '#FF9FF3', '#54A0FF', '#5F27CD']
    
    for idx, (metric_name, scores) in enumerate(metrics_data.items()):
        # 跳过全为0的数据
        if all(score == 0.0 for score in scores):
            continue
            
        plt.plot(steps, scores, marker='o', linewidth=2, 
                label=metric_name.replace('_', ' ').title(), 
                color=colors[idx % len(colors)])
    
    plt.xlabel('Conversation Round', fontsize=12)
    plt.ylabel('Score', fontsize=12)
    plt.title('Evaluate Agent Multi-Dimensional Assessment Trends', fontsize=14, fontweight='bold')
    plt.legend(fontsize=10, bbox_to_anchor=(1.05, 1), loc='upper left')
    plt.grid(True, alpha=0.3)
    plt.ylim(0, 5.5)
    
    # 添加数值标签
    for metric_name, scores in metrics_data.items():
        if not all(score == 0.0 for score in scores):
            for i, score in enumerate(scores):
                if score > 0:
                    plt.annotate(f'{score:.1f}', (steps[i], score), 
                               textcoords="offset points", 
                               xytext=(0, 5), ha='center', fontsize=8)
    
    plt.tight_layout()
    plt.savefig(os.path.join(output_dir, 'evaluate_metrics_trends.png'), dpi=300, bbox_inches='tight')
    plt.close()
    
    # 绘制子图
    _, axes = plt.subplots(2, 4, figsize=(20, 12))
    axes = axes.flatten()
    
    for idx, (metric_name, scores) in enumerate(metrics_data.items()):
        if idx >= len(axes):
            break
            
        ax = axes[idx]
        if not all(score == 0.0 for score in scores):
            ax.plot(steps, scores, marker='o', linewidth=2, color=colors[idx])
            ax.set_title(metric_name.replace('_', ' ').title(), fontsize=12)
            ax.set_xlabel('Conversation Round')
            ax.set_ylabel('Score')
            ax.grid(True, alpha=0.3)
            ax.set_ylim(0, 5.5)
        else:
            ax.text(0.5, 0.5, 'No Data', ha='center', va='center', transform=ax.transAxes)
    
    # 隐藏多余的子图
    for idx in range(len(metrics_data), len(axes)):
        axes[idx].set_visible(False)
    
    plt.tight_layout()
    plt.savefig(os.path.join(output_dir, 'evaluate_metrics_subplots.png'), dpi=300, bbox_inches='tight')
    plt.close()


def save_metrics_data(metrics_data: Dict[str, List[float]], output_dir: str):
    """
    保存评估指标数据到JSON文件
    
    Args:
        metrics_data: 各维度指标数据
        output_dir: 输出目录
    """
    # 转换为更易读的格式
    formatted_data = {
        '维度': list(metrics_data.keys()),
        '步骤': list(range(1, len(next(iter(metrics_data.values()))) + 1)),
        '各维度得分': {}
    }
    
    for metric, scores in metrics_data.items():
        formatted_data['各维度得分'][metric] = scores
    
    with open(os.path.join(output_dir, 'evaluate_metrics_data.json'), 'w', encoding='utf-8') as f:
        json.dump(formatted_data, f, ensure_ascii=False, indent=2)
    
    # 保存简化格式
    simplified_data = {
        '轮次': list(range(1, len(next(iter(metrics_data.values()))) + 1))
    }
    simplified_data.update(metrics_data)
    
    with open(os.path.join(output_dir, 'evaluate_metrics_summary.json'), 'w', encoding='utf-8') as f:
        json.dump(simplified_data, f, ensure_ascii=False, indent=2)


def generate_report(metrics_data: Dict[str, List[float]], output_dir: str):
    """
    生成评估报告
    
    Args:
        metrics_data: 各维度指标数据
        output_dir: 输出目录
    """
    report_lines = [
        "# Evaluate Agent Assessment Report",
        "",
        "## Average Scores by Dimension",
        ""
    ]
    
    for metric_name, scores in metrics_data.items():
        valid_scores = [s for s in scores if s > 0]
        if valid_scores:
            avg_score = np.mean(valid_scores)
            max_score = max(valid_scores)
            min_score = min(valid_scores)
            report_lines.append(
                f"- **{metric_name.replace('_', ' ').title()}**: 平均 {avg_score:.2f} (最高: {max_score:.2f}, 最低: {min_score:.2f})"
            )
    
    report_lines.extend([
        "",
        "",
        "## 分析",
        "",
        "### 表现良好的维度 (平均得分>4.0):"
    ])
    
    good_metrics = []
    for metric_name, scores in metrics_data.items():
        valid_scores = [s for s in scores if s > 0]
        if valid_scores and np.mean(valid_scores) > 4.0:
            good_metrics.append(metric_name.replace('_', ' ').title())
    
    if good_metrics:
        report_lines.extend([f"- {metric}" for metric in good_metrics])
    else:
        report_lines.append("- 无")
    
    report_lines.extend([
        "",
        "### 需要改进的维度（平均得分<2.0）:"
    ])
    
    poor_metrics = []
    for metric_name, scores in metrics_data.items():
        valid_scores = [s for s in scores if s > 0]
        if valid_scores and np.mean(valid_scores) < 2.0:
            poor_metrics.append(metric_name.replace('_', ' ').title())
    
    if poor_metrics:
        report_lines.extend([f"- {metric}" for metric in poor_metrics])
    else:
        report_lines.append("- 无")
    
    with open(os.path.join(output_dir, 'evaluate_report.md'), 'w', encoding='utf-8') as f:
        f.write('\n'.join(report_lines))


def main():
    """主函数"""
    import sys
    
    # 从命令行参数获取路径，如果没有提供则使用默认值
    if len(sys.argv) >= 3:
        data_dir = Path(sys.argv[1])
        output_dir = Path(sys.argv[2])
    else:
        base_dir = Path(__file__).parent.parent
        data_dir = base_dir / "results" / "results0902"
        output_dir = base_dir / "analysis" / "0902"
    
    # 创建输出目录
    output_dir.mkdir(parents=True, exist_ok=True)
    
    print(f"正在加载数据从: {data_dir}")
    workflow_data = load_workflow_data(str(data_dir), str(output_dir), limit=5000)
    print(f"成功加载 {len(workflow_data)} 个病例数据")
    
    if not workflow_data:
        print("未找到有效的工作流数据")
        return
    
    print("正在计算评估指标...")
    metrics_data = calculate_metrics_by_step(workflow_data)
    
    print("评估维度统计结果:")
    for metric, scores in metrics_data.items():
        valid_scores = [s for s in scores if s > 0]
        if valid_scores:
            avg_score = np.mean(valid_scores)
            print(f"  {metric}: 平均 {avg_score:.2f} (轮次: {len(valid_scores)})")
    
    print("正在生成图表...")
    plot_metrics_curves(metrics_data, str(output_dir))
    
    print("正在保存数据...")
    save_metrics_data(metrics_data, str(output_dir))
    
    print("正在生成报告...")
    generate_report(metrics_data, str(output_dir))
    
    print(f"分析完成！结果已保存到: {output_dir}")
    print("输出文件:")
    print("  - evaluate_metrics_data.json: 详细数据")
    print("  - evaluate_metrics_summary.json: 简化数据")
    print("  - evaluate_metrics_trends.png: 趋势图")
    print("  - evaluate_metrics_subplots.png: 子图")
    print("  - evaluate_report.md: 评估报告")


if __name__ == "__main__":
    main()