AIclinicalresearch/backend/src/modules/dc/tool-b/services/DualModelExtractionService.ts

/**
 * DC模块 - 双模型提取服务
 * 
 * 功能：
 * - 并发调用DeepSeek-V3和Qwen-Max进行文本提取
 * - PII脱敏处理
 * - JSON解析与容错
 * - Token统计
 * - 异步任务管理
 * 
 * 平台能力复用：
 * - ✅ LLMFactory: LLM调用
 * - ✅ jobQueue: 异步任务
 * - ✅ logger: 日志记录
 * - ✅ prisma: 数据库操作
 */

import { LLMFactory } from '../../../../common/llm/adapters/LLMFactory.js';
import { logger } from '../../../../common/logging/index.js';
import { prisma } from '../../../../config/database.js';

export interface ExtractionInput {
  text: string;
  fields: { name: string; desc: string }[];
  promptTemplate: string;
}

export interface ExtractionOutput {
  result: Record<string, string>;
  tokensUsed: number;
  rawOutput: any;
}

export class DualModelExtractionService {
  /**
   * 双模型并发提取
   * 
   * @param input 提取输入
   * @param taskId 任务ID
   * @param itemId 记录ID
   * @returns 双模型结果
   */
  async extract(input: ExtractionInput, taskId: string, itemId: string): Promise<{
    resultA: ExtractionOutput;
    resultB: ExtractionOutput;
  }> {
    try {
      logger.info('[DualExtraction] Starting extraction', { taskId, itemId });
      
      // 1. PII脱敏
      const maskedText = this.maskPII(input.text);
      
      // 2. 构建Prompt
      const prompt = this.buildPrompt(maskedText, input.fields, input.promptTemplate);
      
      // 3. 并发调用两个模型（DeepSeek & Qwen）
      const [resultA, resultB] = await Promise.allSettled([
        this.callModel('deepseek', prompt, input.fields),
        this.callModel('qwen', prompt, input.fields)
      ]);
      
      // 4. 处理结果
      if (resultA.status === 'rejected' || resultB.status === 'rejected') {
        logger.error('[DualExtraction] One or both models failed', {
          taskId,
          itemId,
          errorA: resultA.status === 'rejected' ? resultA.reason : null,
          errorB: resultB.status === 'rejected' ? resultB.reason : null
        });
        throw new Error('Dual model extraction failed');
      }
      
      logger.info('[DualExtraction] Extraction completed', {
        taskId,
        itemId,
        tokensA: resultA.value.tokensUsed,
        tokensB: resultB.value.tokensUsed
      });
      
      return {
        resultA: resultA.value,
        resultB: resultB.value
      };
      
    } catch (error) {
      logger.error('[DualExtraction] Extraction failed', { error, taskId, itemId });
      throw error;
    }
  }
  
  /**
   * PII脱敏
   * 
   * 使用正则表达式替换敏感信息：
   * - 姓名：张**
   * - 身份证号：3301********1234
   * - 手机号：138****5678
   */
  private maskPII(text: string): string {
    let masked = text;
    
    // 手机号脱敏：138****5678
    masked = masked.replace(/1[3-9]\d{9}/g, (match) => {
      return match.substring(0, 3) + '****' + match.substring(7);
    });
    
    // 身份证号脱敏：330102********1234
    masked = masked.replace(/\d{6}(19|20)\d{2}(0[1-9]|1[0-2])(0[1-9]|[12]\d|3[01])\d{3}[\dxX]/g, (match) => {
      return match.substring(0, 6) + '********' + match.substring(14);
    });
    
    // 简单的姓名脱敏（匹配：患者xxx、姓名：xxx）
    masked = masked.replace(/(患者|姓名[:：])\s*([^\s，。,]{2,4})/g, (match, prefix, name) => {
      if (name.length === 2) {
        return prefix + name[0] + '*';
      }
      return prefix + name[0] + '*'.repeat(name.length - 1);
    });
    
    return masked;
  }
  
  /**
   * 构建Prompt
   */
  private buildPrompt(text: string, fields: { name: string; desc: string }[], template: string): string {
    // 在模板末尾添加病历文本
    return `${template}

**病历原文：**
${text}

请严格按照JSON格式输出，不要有任何额外文字。`;
  }
  
  /**
   * 调用单个模型
   */
  private async callModel(
    modelType: 'deepseek' | 'qwen',
    prompt: string,
    fields: { name: string; desc: string }[]
  ): Promise<ExtractionOutput> {
    try {
      // 🔑 使用LLMFactory获取适配器（正确的方法）
      const modelName = modelType === 'deepseek' ? 'deepseek-v3' : 'qwen3-72b';
      
      logger.info(`[${modelType.toUpperCase()}] Getting adapter`, { modelName });
      const adapter = LLMFactory.getAdapter(modelName as any);
      logger.info(`[${modelType.toUpperCase()}] Adapter created successfully`);
      
      logger.info(`[${modelType.toUpperCase()}] Calling model with prompt`, { 
        modelName,
        promptLength: prompt.length,
        promptPreview: prompt.substring(0, 100) + '...'
      });
      
      // 🔑 调用LLM（使用chat方法，符合ILLMAdapter接口）
      const startTime = Date.now();
      const response = await adapter.chat([
        { role: 'user', content: prompt }
      ], {
        temperature: 0, // 最大确定性
        maxTokens: 1000
      });
      const elapsedTime = Date.now() - startTime;
      
      logger.info(`[${modelType.toUpperCase()}] Model responded successfully`, {
        modelName,
        tokensUsed: response.usage?.totalTokens,
        elapsedMs: elapsedTime,
        contentLength: response.content.length,
        contentPreview: response.content.substring(0, 200)
      });
      
      // 解析JSON（3层容错）
      logger.info(`[${modelType.toUpperCase()}] Parsing JSON response`);
      const result = this.parseJSON(response.content, fields);
      logger.info(`[${modelType.toUpperCase()}] JSON parsed successfully`, { 
        fieldCount: Object.keys(result).length 
      });
      
      return {
        result,
        tokensUsed: response.usage?.totalTokens || 0,
        rawOutput: response.content
      };
      
    } catch (error: any) {
      logger.error(`[${modelType.toUpperCase()}] Model call failed`, { 
        error: error.message,
        stack: error.stack,
        modelType 
      });
      throw error;
    }
  }
  
  /**
   * 解析JSON（3层容错策略）
   * 
   * 1. 直接JSON.parse
   * 2. 提取```json代码块
   * 3. 提取{}内容
   */
  private parseJSON(text: string, fields: { name: string; desc: string }[]): Record<string, string> {
    // 策略1：直接解析
    try {
      const parsed = JSON.parse(text);
      if (this.validateFields(parsed, fields)) {
        return parsed;
      }
    } catch (e) {
      // 继续下一个策略
    }
    
    // 策略2：提取```json代码块
    const codeBlockMatch = text.match(/```json\s*\n([\s\S]*?)\n```/);
    if (codeBlockMatch) {
      try {
        const parsed = JSON.parse(codeBlockMatch[1]);
        if (this.validateFields(parsed, fields)) {
          return parsed;
        }
      } catch (e) {
        // 继续下一个策略
      }
    }
    
    // 策略3：提取第一个完整的{}对象
    const objectMatch = text.match(/\{[\s\S]*\}/);
    if (objectMatch) {
      try {
        const parsed = JSON.parse(objectMatch[0]);
        if (this.validateFields(parsed, fields)) {
          return parsed;
        }
      } catch (e) {
        // 解析失败
      }
    }
    
    // 所有策略失败，返回空对象
    logger.warn('[JSON] All parse strategies failed', { text });
    const emptyResult: Record<string, string> = {};
    fields.forEach(f => {
      emptyResult[f.name] = '解析失败';
    });
    return emptyResult;
  }
  
  /**
   * 验证字段完整性
   */
  private validateFields(parsed: any, fields: { name: string; desc: string }[]): boolean {
    if (!parsed || typeof parsed !== 'object') {
      return false;
    }
    
    // 检查所有必需字段是否存在
    return fields.every(f => parsed.hasOwnProperty(f.name));
  }
  
  /**
   * 批量提取（异步任务）
   * 
   * @param taskId 任务ID
   */
  async batchExtract(taskId: string): Promise<void> {
    try {
      logger.info('[Batch] ===== Starting batch extraction =====', { taskId });
      
      // 1. 获取任务
      logger.info('[Batch] Step 1: Fetching task from database', { taskId });
      const task = await prisma.dCExtractionTask.findUnique({
        where: { id: taskId },
        include: { items: true }
      });
      
      if (!task) {
        logger.error('[Batch] Task not found in database', { taskId });
        throw new Error(`Task not found: ${taskId}`);
      }
      
      logger.info('[Batch] Task fetched successfully', { 
        taskId, 
        itemCount: task.items.length,
        diseaseType: task.diseaseType,
        reportType: task.reportType
      });
      
      // 2. 更新任务状态
      await prisma.dCExtractionTask.update({
        where: { id: taskId },
        data: {
          status: 'processing',
          startedAt: new Date()
        }
      });
      
      // 3. 获取模板
      const template = await prisma.dCTemplate.findUnique({
        where: {
          diseaseType_reportType: {
            diseaseType: task.diseaseType,
            reportType: task.reportType
          }
        }
      });
      
      if (!template) {
        throw new Error(`Template not found: ${task.diseaseType}/${task.reportType}`);
      }
      
      const fields = template.fields as { name: string; desc: string }[];
      
      // 4. 逐条处理
      let processedCount = 0;
      let cleanCount = 0;
      let conflictCount = 0;
      let totalTokens = 0;
      
      for (const item of task.items) {
        try {
          // 双模型提取
          const { resultA, resultB } = await this.extract(
            {
              text: item.originalText,
              fields,
              promptTemplate: template.promptTemplate
            },
            taskId,
            item.id
          );
          
          // 检测冲突（由ConflictDetectionService处理，这里暂时简单比较）
          const hasConflict = JSON.stringify(resultA.result) !== JSON.stringify(resultB.result);
          
          // 更新记录
          await prisma.dCExtractionItem.update({
            where: { id: item.id },
            data: {
              resultA: resultA.result as any,
              resultB: resultB.result as any,
              tokensA: resultA.tokensUsed,
              tokensB: resultB.tokensUsed,
              status: hasConflict ? 'conflict' : 'clean',
              finalResult: (hasConflict ? null : resultA.result) as any // 一致时自动采纳
            }
          });
          
          processedCount++;
          if (hasConflict) {
            conflictCount++;
          } else {
            cleanCount++;
          }
          totalTokens += resultA.tokensUsed + resultB.tokensUsed;
          
          // 更新任务进度
          await prisma.dCExtractionTask.update({
            where: { id: taskId },
            data: {
              processedCount,
              cleanCount,
              conflictCount,
              totalTokens
            }
          });
          
        } catch (error) {
          logger.error('[Batch] Item extraction failed', { error, itemId: item.id });
          
          await prisma.dCExtractionItem.update({
            where: { id: item.id },
            data: {
              status: 'failed',
              error: String(error)
            }
          });
        }
      }
      
      // 5. 完成任务
      await prisma.dCExtractionTask.update({
        where: { id: taskId },
        data: {
          status: 'completed',
          completedAt: new Date()
        }
      });
      
      logger.info('[Batch] Batch extraction completed', {
        taskId,
        processedCount,
        cleanCount,
        conflictCount,
        totalTokens
      });
      
    } catch (error) {
      logger.error('[Batch] Batch extraction failed', { error, taskId });
      
      // 更新任务为失败状态
      await prisma.dCExtractionTask.update({
        where: { id: taskId },
        data: {
          status: 'failed',
          error: String(error)
        }
      });
      
      throw error;
    }
  }
}

// 导出单例
export const dualModelExtractionService = new DualModelExtractionService();