AIclinicalresearch/backend/test-tool-c-day3.mjs

/**
 * Tool C Day 3 测试脚本
 * 
 * 测试内容：
 * 1. 10个Few-shot示例场景测试
 * 2. AI自我修正机制测试（重试）
 * 3. 端到端测试
 * 
 * 前提：
 * - 需要先创建一个Session（使用Day 2的upload接口）
 * - 需要Python服务运行（端口8000）
 * - 需要后端服务运行（端口3000）
 * 
 * 执行方式：node test-tool-c-day3.mjs
 */

import axios from 'axios';
import FormData from 'form-data';
import * as XLSX from 'xlsx';

const BASE_URL = 'http://localhost:3000';
const API_PREFIX = '/api/v1/dc/tool-c';

let testSessionId = null;

// ==================== 辅助函数 ====================

function printSection(title) {
  console.log('\n' + '='.repeat(70));
  console.log(`  ${title}`);
  console.log('='.repeat(70) + '\n');
}

function printSuccess(message) {
  console.log('✅ ' + message);
}

function printError(message) {
  console.log('❌ ' + message);
}

function printInfo(message) {
  console.log('ℹ️  ' + message);
}

// ==================== 准备测试Session ====================

async function createTestSession() {
  printSection('准备：创建测试Session');
  
  try {
    // 创建测试Excel数据
    const testData = [
      { patient_id: 'P001', name: '张三', age: 25, gender: '男', diagnosis: '感冒', sbp: 120, dbp: 80, weight: 70, height: 175, BMI: '', creatinine: '>100', check_date: '2024-01-01' },
      { patient_id: 'P002', name: '李四', age: 65, gender: '女', diagnosis: '高血压', sbp: 150, dbp: 95, weight: 65, height: 160, BMI: '', creatinine: '<0.1', check_date: '2024-01-05' },
      { patient_id: 'P003', name: '王五', age: 45, gender: '男', diagnosis: '糖尿病', sbp: 135, dbp: 85, weight: 80, height: 170, BMI: '', creatinine: '85', check_date: '2024-01-03' },
      { patient_id: 'P004', name: '赵六', age: 70, gender: '女', diagnosis: '冠心病', sbp: 160, dbp: 100, weight: 60, height: 155, BMI: '', creatinine: '120', check_date: '2024-01-10' },
      { patient_id: 'P005', name: '钱七', age: 35, gender: '男', diagnosis: '胃炎', sbp: 110, dbp: 70, weight: 75, height: 180, BMI: '', creatinine: '-', check_date: '2024-01-08' },
      { patient_id: 'P003', name: '王五', age: 45, gender: '男', diagnosis: '糖尿病', sbp: 138, dbp: 88, weight: 82, height: 170, BMI: '', creatinine: '88', check_date: '2024-01-12' },  // 重复ID，日期更新
    ];
    
    // 生成Excel
    const ws = XLSX.utils.json_to_sheet(testData);
    const wb = XLSX.utils.book_new();
    XLSX.utils.book_append_sheet(wb, ws, 'Sheet1');
    const excelBuffer = XLSX.write(wb, { type: 'buffer', bookType: 'xlsx' });
    
    // 上传创建Session
    const form = new FormData();
    form.append('file', excelBuffer, {
      filename: 'test-medical-data-day3.xlsx',
      contentType: 'application/vnd.openxmlformats-officedocument.spreadsheetml.sheet',
    });
    
    const response = await axios.post(
      `${BASE_URL}${API_PREFIX}/sessions/upload`,
      form,
      { headers: form.getHeaders(), timeout: 10000 }
    );
    
    if (response.data.success) {
      testSessionId = response.data.data.sessionId;
      printSuccess(`Session创建成功: ${testSessionId}`);
      printInfo(`数据: ${testData.length}行 x ${Object.keys(testData[0]).length}列`);
      return true;
    } else {
      printError('Session创建失败');
      return false;
    }
  } catch (error) {
    printError('Session创建异常: ' + error.message);
    return false;
  }
}

// ==================== AI测试函数 ====================

async function testAIGenerate(testName, userMessage, shouldSucceed = true) {
  printSection(`测试: ${testName}`);
  
  try {
    printInfo(`用户需求: ${userMessage}`);
    
    const response = await axios.post(
      `${BASE_URL}${API_PREFIX}/ai/generate`,
      {
        sessionId: testSessionId,
        message: userMessage
      },
      { timeout: 30000 }  // AI调用可能需要较长时间
    );
    
    if (response.data.success) {
      printSuccess('AI生成代码成功');
      console.log('\n生成的代码:');
      console.log('```python');
      console.log(response.data.data.code);
      console.log('```\n');
      console.log('解释:', response.data.data.explanation);
      console.log('MessageID:', response.data.data.messageId);
      return { success: true, data: response.data.data };
    } else {
      if (shouldSucceed) {
        printError('AI生成失败: ' + response.data.error);
      } else {
        printInfo('预期失败: ' + response.data.error);
      }
      return { success: false, error: response.data.error };
    }
  } catch (error) {
    printError('AI生成异常: ' + (error.response?.data?.error || error.message));
    return { success: false, error: error.message };
  }
}

async function testAIProcess(testName, userMessage) {
  printSection(`测试（一步到位）: ${testName}`);
  
  try {
    printInfo(`用户需求: ${userMessage}`);
    
    const response = await axios.post(
      `${BASE_URL}${API_PREFIX}/ai/process`,
      {
        sessionId: testSessionId,
        message: userMessage,
        maxRetries: 3
      },
      { timeout: 60000 }  // 带重试可能需要更长时间
    );
    
    if (response.data.success) {
      printSuccess(`执行成功${response.data.data.retryCount > 0 ? `（重试${response.data.data.retryCount}次）` : ''}`);
      console.log('\n生成的代码:');
      console.log('```python');
      console.log(response.data.data.code);
      console.log('```\n');
      console.log('解释:', response.data.data.explanation);
      
      if (response.data.data.executeResult.success) {
        printSuccess('代码执行成功');
        console.log('数据预览（前5行）:');
        console.log(JSON.stringify(response.data.data.executeResult.newDataPreview?.slice(0, 5), null, 2));
      }
      
      return { success: true, data: response.data.data };
    } else {
      printError('处理失败: ' + response.data.error);
      return { success: false, error: response.data.error };
    }
  } catch (error) {
    printError('处理异常: ' + (error.response?.data?.error || error.message));
    return { success: false, error: error.message };
  }
}

// ==================== 主测试函数 ====================

async function runAllTests() {
  console.log('\n' + '🚀'.repeat(35));
  console.log('   Tool C Day 3 测试 - AI代码生成');
  console.log('🚀'.repeat(35));
  
  const results = {};
  
  try {
    // 0. 准备测试Session
    const sessionCreated = await createTestSession();
    if (!sessionCreated) {
      printError('测试Session创建失败，无法继续');
      return;
    }
    
    await new Promise(resolve => setTimeout(resolve, 2000));
    
    // ==================== 10个Few-shot示例测试 ====================
    
    // 测试1: 统一缺失值标记
    let result = await testAIProcess(
      '示例1: 统一缺失值标记',
      '把所有代表缺失的符号（-、不详、NA、N/A）统一替换为标准空值'
    );
    results['示例1-缺失值'] = result.success;
    await new Promise(resolve => setTimeout(resolve, 3000));
    
    // 测试2: 数值列清洗
    result = await testAIProcess(
      '示例2: 数值列清洗',
      '把creatinine列里的非数字符号去掉，<0.1按0.05处理，转为数值类型'
    );
    results['示例2-数值清洗'] = result.success;
    await new Promise(resolve => setTimeout(resolve, 3000));
    
    // 测试3: 分类变量编码
    result = await testAIProcess(
      '示例3: 分类变量编码',
      '把gender列转为数字，男=1，女=0'
    );
    results['示例3-编码'] = result.success;
    await new Promise(resolve => setTimeout(resolve, 3000));
    
    // 测试4: 连续变量分箱
    result = await testAIProcess(
      '示例4: 连续变量分箱',
      '把age列按18岁、60岁分为未成年、成年、老年三组'
    );
    results['示例4-分箱'] = result.success;
    await new Promise(resolve => setTimeout(resolve, 3000));
    
    // 测试5: BMI计算
    result = await testAIProcess(
      '示例5: BMI计算',
      '根据weight和height计算BMI，并标记BMI≥28为肥胖'
    );
    results['示例5-BMI'] = result.success;
    await new Promise(resolve => setTimeout(resolve, 3000));
    
    // 测试6: 条件筛选
    result = await testAIProcess(
      '示例6: 条件筛选',
      '筛选出年龄≥60岁、且sbp≥140的患者'
    );
    results['示例6-筛选'] = result.success;
    await new Promise(resolve => setTimeout(resolve, 3000));
    
    // 测试7: 智能去重
    result = await testAIProcess(
      '示例7: 智能去重',
      '按patient_id去重，保留check_date最新的记录'
    );
    results['示例7-去重'] = result.success;
    await new Promise(resolve => setTimeout(resolve, 3000));
    
    // 测试8: 中位数填补（简化版，跳过多重插补）
    result = await testAIProcess(
      '示例8: 缺失值填补',
      '用age列的中位数填补age列的缺失值'
    );
    results['示例8-填补'] = result.success;
    await new Promise(resolve => setTimeout(resolve, 3000));
    
    // 测试9: 统计汇总
    result = await testAIProcess(
      '示例9: 统计汇总',
      '按diagnosis分组，统计每个诊断的平均年龄和患者数量'
    );
    results['示例9-统计'] = result.success;
    await new Promise(resolve => setTimeout(resolve, 3000));
    
    // 测试10: 复杂计算
    result = await testAIProcess(
      '示例10: 复杂计算',
      '根据sbp判断血压分类：正常(<140)、高血压I级(140-159)、高血压II级(≥160)'
    );
    results['示例10-分类'] = result.success;
    
    // ==================== 对话历史测试 ====================
    
    printSection('测试: 获取对话历史');
    try {
      const historyResponse = await axios.get(
        `${BASE_URL}${API_PREFIX}/ai/history/${testSessionId}?limit=5`
      );
      
      if (historyResponse.data.success) {
        printSuccess(`获取历史成功: ${historyResponse.data.data.count}条`);
        results['对话历史'] = true;
      } else {
        printError('获取历史失败');
        results['对话历史'] = false;
      }
    } catch (error) {
      printError('获取历史异常: ' + error.message);
      results['对话历史'] = false;
    }
    
  } catch (error) {
    printError('测试过程中发生异常: ' + error.message);
    console.error(error);
  }
  
  // 汇总结果
  printSection('测试结果汇总');
  
  let passed = 0;
  let total = 0;
  
  for (const [testName, result] of Object.entries(results)) {
    total++;
    if (result) {
      passed++;
      console.log(`${testName.padEnd(20)}: ✅ 通过`);
    } else {
      console.log(`${testName.padEnd(20)}: ❌ 失败`);
    }
  }
  
  console.log('\n' + '-'.repeat(70));
  console.log(`总计: ${passed}/${total} 通过 (${((passed/total)*100).toFixed(1)}%)`);
  console.log('-'.repeat(70));
  
  if (passed === total) {
    console.log('\n🎉 所有测试通过！Day 3 AI功能完成！\n');
  } else if (passed >= total * 0.7) {
    console.log(`\n⚠️  有 ${total - passed} 个测试失败，但通过率≥70%，基本可用\n`);
  } else {
    console.log(`\n❌  通过率过低，需要调试\n`);
  }
}

// 执行测试
runAllTests()
  .then(() => {
    console.log('测试完成');
    process.exit(0);
  })
  .catch((error) => {
    console.error('测试失败:', error);
    process.exit(1);
  });