/**
 * 卒中数据测试 - 国际模型对比
 * 
 * 目的：对比国内模型（DeepSeek+Qwen）vs 国际模型（GPT-4o+Claude）
 * 
 * 测试假设：
 * 1. 如果国际模型准确率更高 → 是模型能力问题
 * 2. 如果国际模型准确率相似 → 是Prompt或理解差异问题
 */

import * as fs from 'fs';
import * as path from 'path';
import * as XLSX from 'xlsx';
import { fileURLToPath } from 'url';
import { llmScreeningService } from '../src/modules/asl/services/llmScreeningService.js';

const __filename = fileURLToPath(import.meta.url);
const __dirname = path.dirname(__filename);

// ========================================
// 📋 1. 读取PICOS和标准
// ========================================

console.log('📖 正在读取PICOS和纳排标准...\n');

const picosPath = path.join(
  __dirname,
  '../../docs/03-业务模块/ASL-AI智能文献/05-测试文档/03-测试数据/screening/测试案例的PICOS、纳入标准、排除标准.txt'
);

const picosContent = fs.readFileSync(picosPath, 'utf-8');

// 解析PICOS（简化版）
const picoCriteria = {
  population: '非心源性缺血性卒中患者、亚洲人群',
  intervention: '抗血小板药物/抗凝药物/溶栓药物（阿司匹林、氯吡格雷、替格瑞洛、达比加群等）',
  comparison: '安慰剂或常规治疗',
  outcome: '卒中进展、复发、残疾程度、死亡率、出血事件等',
  studyDesign: 'SR、RCT、RWE、OBS'
};

const inclusionCriteria = `
1. 研究对象为非心源性缺血性卒中患者
2. 研究人群为亚洲人群（优先）
3. 干预措施为抗血小板/抗凝/溶栓药物
4. 对照组为安慰剂或常规治疗
5. 研究时间在2020年之后
6. 研究设计为SR、RCT、RWE、OBS
`;

const exclusionCriteria = `
1. 综述、病例报告、会议摘要
2. 动物实验、体外实验
3. 研究人群非亚洲人群（除非有特殊价值）
4. 研究时间在2020年之前
5. 心源性卒中或出血性卒中
`;

console.log('✅ PICOS标准已加载\n');

// ========================================
// 📋 2. 读取测试案例
// ========================================

console.log('📖 正在读取测试案例...\n');

const excelPath = path.join(
  __dirname,
  '../../docs/03-业务模块/ASL-AI智能文献/05-测试文档/03-测试数据/screening/Test Cases.xlsx'
);

const workbook = XLSX.read(fs.readFileSync(excelPath), { type: 'buffer' });
const sheetName = workbook.SheetNames[0];
const worksheet = workbook.Sheets[sheetName];
const data = XLSX.utils.sheet_to_json(worksheet);

console.log(`✅ 读取到 ${data.length} 条数据\n`);

// 选择测试样本：2个Included + 3个Excluded
const includedCases = data.filter((row: any) => 
  row['Decision']?.toString().toLowerCase().includes('include')
).slice(0, 2);

const excludedCases = data.filter((row: any) => 
  row['Decision']?.toString().toLowerCase().includes('exclude')
).slice(0, 3);

const testCases = [...includedCases, ...excludedCases];

console.log(`✅ 选择测试样本: ${testCases.length}篇（2 Included + 3 Excluded）\n`);

// ========================================
// 🧪 3. 定义测试模型组合
// ========================================

const modelPairs = [
  {
    name: '国内模型组合',
    model1: 'deepseek-chat',
    model2: 'qwen3-72b',
    description: 'DeepSeek-V3 + Qwen3-Max（当前使用）'
  },
  {
    name: '国际模型组合',
    model1: 'gpt-4o',
    model2: 'claude-sonnet-4.5',
    description: 'GPT-4o + Claude-4.5（国际顶级模型）'
  }
];

// ========================================
// 🧪 4. 执行测试
// ========================================

interface TestResult {
  caseIndex: number;
  title: string;
  humanDecision: string;
  aiDecision: string;
  model1Result: any;
  model2Result: any;
  isCorrect: boolean;
  hasConflict: boolean;
  processingTime: number;
}

async function testModelPair(
  pairName: string,
  model1: string,
  model2: string,
  cases: any[]
): Promise<TestResult[]> {
  console.log(`\n${'='.repeat(60)}`);
  console.log(`🧪 测试模型组合: ${pairName}`);
  console.log(`${'='.repeat(60)}\n`);

  const results: TestResult[] = [];

  for (let i = 0; i < cases.length; i++) {
    const testCase = cases[i];
    const title = testCase['title'] || '';
    const abstract = testCase['abstract'] || '';
    const humanDecision = testCase['Decision'] || '';

    console.log(`\n[${i + 1}/${cases.length}] 正在筛选...`);
    console.log(`标题: ${title.substring(0, 60)}...`);
    console.log(`人类决策: ${humanDecision}`);

    const startTime = Date.now();

    try {
      const screeningResult = await llmScreeningService.dualModelScreening(
        `test-case-${i + 1}`,  // literatureId
        title,
        abstract,
        picoCriteria,
        inclusionCriteria,
        exclusionCriteria,
        [model1, model2],  // models参数应该是一个数组
        'standard'  // style参数
      );

      const processingTime = Date.now() - startTime;

      // 标准化决策
      const normalizedHuman = humanDecision.toLowerCase().includes('include') ? 'include' : 'exclude';
      const normalizedAI = screeningResult.finalDecision === 'pending' ? 'uncertain' : screeningResult.finalDecision;
      
      const isCorrect = normalizedAI === normalizedHuman;

      console.log(`AI决策: ${screeningResult.finalDecision} ${isCorrect ? '✅' : '❌'}`);
      console.log(`模型一致: ${!screeningResult.hasConflict ? '✅' : '❌'}`);
      console.log(`处理时间: ${(processingTime / 1000).toFixed(2)}秒`);

      results.push({
        caseIndex: i + 1,
        title: title.substring(0, 100),
        humanDecision: normalizedHuman,
        aiDecision: normalizedAI,
        model1Result: screeningResult.model1Result,
        model2Result: screeningResult.model2Result,
        isCorrect,
        hasConflict: screeningResult.hasConflict,
        processingTime
      });

    } catch (error: any) {
      console.error(`❌ 筛选失败: ${error.message}`);
      results.push({
        caseIndex: i + 1,
        title: title.substring(0, 100),
        humanDecision: humanDecision.toLowerCase().includes('include') ? 'include' : 'exclude',
        aiDecision: 'error',
        model1Result: null,
        model2Result: null,
        isCorrect: false,
        hasConflict: false,
        processingTime: Date.now() - startTime
      });
    }
  }

  return results;
}

// ========================================
// 📊 5. 生成对比报告
// ========================================

function generateComparisonReport(
  domesticResults: TestResult[],
  internationalResults: TestResult[]
) {
  console.log(`\n${'='.repeat(80)}`);
  console.log(`📊 国内 vs 国际模型对比报告`);
  console.log(`${'='.repeat(80)}\n`);

  // 计算指标
  function calculateMetrics(results: TestResult[]) {
    const total = results.length;
    const correct = results.filter(r => r.isCorrect).length;
    const consistent = results.filter(r => !r.hasConflict).length;
    const avgTime = results.reduce((sum, r) => sum + r.processingTime, 0) / total;

    return {
      accuracy: (correct / total * 100).toFixed(1),
      consistency: (consistent / total * 100).toFixed(1),
      avgTime: (avgTime / 1000).toFixed(2),
      correct,
      total
    };
  }

  const domesticMetrics = calculateMetrics(domesticResults);
  const internationalMetrics = calculateMetrics(internationalResults);

  // 对比表格
  console.log('| 指标 | 国内模型 | 国际模型 | 差异 |');
  console.log('|------|----------|----------|------|');
  console.log(`| 准确率 | ${domesticMetrics.accuracy}% (${domesticMetrics.correct}/${domesticMetrics.total}) | ${internationalMetrics.accuracy}% (${internationalMetrics.correct}/${internationalMetrics.total}) | ${(parseFloat(internationalMetrics.accuracy) - parseFloat(domesticMetrics.accuracy)).toFixed(1)}% |`);
  console.log(`| 一致率 | ${domesticMetrics.consistency}% | ${internationalMetrics.consistency}% | ${(parseFloat(internationalMetrics.consistency) - parseFloat(domesticMetrics.consistency)).toFixed(1)}% |`);
  console.log(`| 平均耗时 | ${domesticMetrics.avgTime}秒 | ${internationalMetrics.avgTime}秒 | ${(parseFloat(internationalMetrics.avgTime) - parseFloat(domesticMetrics.avgTime)).toFixed(2)}秒 |`);

  console.log('\n');

  // 逐案例对比
  console.log('📋 逐案例对比:\n');
  for (let i = 0; i < domesticResults.length; i++) {
    const domestic = domesticResults[i];
    const international = internationalResults[i];

    console.log(`[案例 ${i + 1}] ${domestic.title}`);
    console.log(`  人类: ${domestic.humanDecision}`);
    console.log(`  国内模型: ${domestic.aiDecision} ${domestic.isCorrect ? '✅' : '❌'}`);
    console.log(`  国际模型: ${international.aiDecision} ${international.isCorrect ? '✅' : '❌'}`);
    
    if (domestic.aiDecision !== international.aiDecision) {
      console.log(`  ⚠️ 两组模型判断不一致！`);
    }
    console.log('');
  }

  // 结论分析
  console.log('\n' + '='.repeat(80));
  console.log('🎯 结论分析\n');

  const accuracyDiff = parseFloat(internationalMetrics.accuracy) - parseFloat(domesticMetrics.accuracy);

  if (Math.abs(accuracyDiff) <= 10) {
    console.log('✅ 结论: 国内外模型准确率相近（差异≤10%）');
    console.log('   → 问题不在模型能力，而在于：');
    console.log('     1. Prompt设计（可能过于严格）');
    console.log('     2. AI vs 人类对"匹配"的理解差异');
    console.log('     3. 纳排标准本身存在歧义');
    console.log('\n💡 建议: 优化Prompt策略，增加宽松/标准/严格三种模式');
  } else if (accuracyDiff > 10) {
    console.log('✅ 结论: 国际模型显著优于国内模型（差异>10%）');
    console.log('   → 问题在于模型能力差异');
    console.log('   → 国际模型对医学文献的理解更准确');
    console.log('\n💡 建议: 优先使用GPT-4o或Claude-4.5进行筛选');
  } else {
    console.log('✅ 结论: 国内模型优于国际模型（差异>10%）');
    console.log('   → 可能是国内模型对中文医学术语理解更好');
    console.log('   → 或者国内模型更符合中国专家的筛选习惯');
    console.log('\n💡 建议: 继续使用国内模型组合');
  }

  console.log('='.repeat(80) + '\n');

  // 保存详细报告
  const report = {
    testDate: new Date().toISOString(),
    testCases: testCases.length,
    domesticModels: modelPairs[0],
    internationalModels: modelPairs[1],
    domesticMetrics,
    internationalMetrics,
    domesticResults,
    internationalResults,
    conclusion: {
      accuracyDiff,
      analysis: Math.abs(accuracyDiff) <= 10 ? 'Prompt问题' : (accuracyDiff > 10 ? '国际模型更优' : '国内模型更优')
    }
  };

  const reportPath = path.join(__dirname, '../docs/国内外模型对比测试报告.json');
  fs.writeFileSync(reportPath, JSON.stringify(report, null, 2), 'utf-8');
  console.log(`📄 详细报告已保存: ${reportPath}\n`);
}

// ========================================
// 🚀 6. 执行主流程
// ========================================

async function main() {
  console.log('\n🚀 开始国内外模型对比测试\n');
  console.log(`测试样本: ${testCases.length}篇`);
  console.log(`测试组合: 2组`);
  console.log(`预计耗时: ${testCases.length * 2 * 15}秒（约${Math.ceil(testCases.length * 2 * 15 / 60)}分钟）\n`);

  // 测试国内模型
  const domesticResults = await testModelPair(
    modelPairs[0].name,
    modelPairs[0].model1,
    modelPairs[0].model2,
    testCases
  );

  // 等待2秒，避免API限流
  console.log('\n⏳ 等待2秒后测试国际模型...\n');
  await new Promise(resolve => setTimeout(resolve, 2000));

  // 测试国际模型
  const internationalResults = await testModelPair(
    modelPairs[1].name,
    modelPairs[1].model1,
    modelPairs[1].model2,
    testCases
  );

  // 生成对比报告
  generateComparisonReport(domesticResults, internationalResults);

  console.log('✅ 测试完成！\n');
}

main().catch(console.error);