AIclinicalresearch/docs/03-业务模块/ASL-AI智能文献/05-开发记录/2025-11-21-真实LLM集成完成报告.md

真实LLM集æˆ<EFBFBD>完æˆ<EFBFBD>报告

日期: 2025-11-21
任务: å°†Mock AI替æ<C2BF>¢ä¸ºçœŸå®žLLM调用
**状æ€?*: âœ?完æˆ<C3A6>

---

📋 背景

之å‰<EFBFBD>的状æ€?

• âœ?已完æˆ?Prompt 设计（v1.0.0-MVPï¼?
• âœ?已实çŽ?llmScreeningService.ts（真实LLM调用ï¼?
• âœ?已完æˆ<C3A6>测试框架和质é‡<C3A9>验è¯<C3A8>
• â<EFBFBD>?问题: screeningService.ts 中使ç”?mockAIScreening 生æˆ<C3A6>å<EFBFBD>‡æ•°æ<C2B0>?

用户需�

ä»?设置与å<C5BD>¯åŠ?页é<C2B5>¢ä¸Šä¼ çœŸå®žæ–‡çŒ®æ•°æ<C2B0>®å<C2AE>Žï¼Œ**使用真实çš?DeepSeek å’?Qwen API 进行筛é€?*，而ä¸<C3A4>是模拟数æ<C2B0>®ã€?

---

âœ?完æˆ<C3A6>内容

1. 修改 screeningService.ts

文件: backend/src/modules/asl/services/screeningService.ts

核心改动

引入真实LLMæœ<EFBFBD>务:

import { llmScreeningService } from './llmScreeningService.js';

替æ<EFBFBD>¢å¤„ç<EFBFBD>†é€»è¾‘:

// â<>?旧代ç <C3A7>（Mockï¼?
const result = await mockAIScreening(projectId, literature);

// âœ?新代ç <C3A7>（真实LLMï¼?
const screeningResult = await llmScreeningService.dualModelScreening(
  literature.id,
  literature.title,
  literature.abstract,
  picoCriteria,
  inclusionCriteria,
  exclusionCriteria,
  [models[0], models[1]],
  screeningConfig?.style || 'standard',
  literature.authors,
  literature.journal,
  literature.publicationYear
);

新增功能

1. 从项目读å<EFBFBD>–PICOS标准:

   const project = await prisma.aslScreeningProject.findUnique({
     where: { id: projectId },
   });
   
   const picoCriteria = project.picoCriteria;
   const inclusionCriteria = project.inclusionCriteria;
   const exclusionCriteria = project.exclusionCriteria;
   

2. 支æŒ<EFBFBD>自定义模型选择:

   const models = screeningConfig?.models || ['deepseek-chat', 'qwen-max'];
   

3. 详细日志记录:

   logger.info('Processing literature', {
     literatureId: literature.id,
     title: literature.title?.substring(0, 50) + '...',
   });
   

4. 结果映射到数æ<EFBFBD>®åº“æ ¼å¼<EFBFBD>:

   const dbResult = {
     projectId,
     literatureId: literature.id,
     // DeepSeek结果
     dsModelName: screeningResult.deepseekModel,
     dsPJudgment: screeningResult.deepseek.judgment.P,
     // ... 完整的字段映�
   };
   

---

🔄 完整æµ<C3A6>程

用户æ“<EFBFBD>作æµ<EFBFBD>程

1. 访问"设置与å<C5BD>¯åŠ?页é<C2B5>¢
   �
2. 填写 PICOS 标准
   �
3. 上传 Excel 文献列表（例如：199篇）
   �
4. 点击"开始AIåˆ<C3A5>ç­›"
   �
5. å<>Žç«¯è‡ªåŠ¨å¤„ç<E2809E>†ï¼?
   a. 创建项目
   b. 导入文献
   c. å<>¯åŠ¨ç­›é€‰ä»»åŠ?
   �
6. 真实LLM处ç<E2809E>†ï¼ˆæ¯<C3A6>篇约10-15秒）
   a. 调用 DeepSeek API
   b. 调用 Qwen API
   c. 对比结果，检测冲�
   d. ä¿<C3A4>存到数æ<C2B0>®åº“
   �
7. å‰<C3A5>端自动跳转åˆ?审核工作å<C593>?
   �
8. 显示真实的AI筛选结�

技术æµ<EFBFBD>ç¨?

å‰<EFBFBD>端: TitleScreeningSettings.tsx
  �POST /api/v1/asl/literatures/import
  
å<>Žç«¯: literatureController.ts
  �importLiteratures()
  �startScreeningTask()
  
å<>Žç«¯: screeningService.ts
  �processLiteraturesInBackground()
  �for each literature:
      �llmScreeningService.dualModelScreening()
      
å<>Žç«¯: llmScreeningService.ts
  �Promise.all([
      screenWithModel('deepseek-chat', ...),
      screenWithModel('qwen-max', ...),
    ])
  
å<>Žç«¯: LLMFactory
  �getAdapter('deepseek-v3')
  �getAdapter('qwen3-72b')
  
真实API调用
  �DeepSeek API
  �Qwen API
  
结果ä¿<C3A4>å­˜
  �AslScreeningResult �
  
å‰<C3A5>端: ScreeningWorkbench.tsx
  �GET /api/v1/asl/projects/:projectId/screening-results
  �显示真实结果

---

â<EFBFBD>±ï¸<EFBFBD> 性能预期

å<EFBFBD>•ç¯‡æ–‡çŒ®å¤„ç<EFBFBD>†æ—¶é—´

步骤	耗时（串行）
DeepSeek API 调用	5-10�
Qwen API 调用	5-10�
结果ä¿<EFBFBD>å­˜	0.1ç§?
总计	*10-20�

批é‡<EFBFBD>处ç<EFBFBD>†æ—¶é—´ï¼?99篇）

模å¼<EFBFBD>	耗时	说明
串行处ç<EFBFBD>†	33-66分钟	当å‰<EFBFBD>实现（é<EFBFBD>¿å…<EFBFBD>APIé™<EFBFBD>æµ<EFBFBD>）
并å<EFBFBD>‘处ç<EFBFBD>†ï¼?个）	11-22分钟	å<EFBFBD>¯é€‰ä¼˜åŒ–（需测试ï¼?
并å<EFBFBD>‘处ç<EFBFBD>†ï¼?0个）	3-7分钟	风险：å<EFBFBD>¯èƒ½è§¦å<EFBFBD>‘APIé™<EFBFBD>é¢<EFBFBD>

当å‰<EFBFBD>ç­–ç•¥: 串行处ç<E2809E>†ï¼ˆç¨³å®šä¼˜å…ˆï¼‰

---

🎯 与Mockæ•°æ<C2B0>®çš„对æ¯?

Mock æ•°æ<C2B0>®ï¼ˆæ—§ï¼?

// â<>?å<>‡æ•°æ<C2B0>?
dsPEvidence: "模拟è¯<C3A8>æ<EFBFBD>®: 研究人群与PICO中的P标准匹é…<C3A9>"
dsReason: "基于标题和摘è¦<C3A8>分æž<C3A6>，该文献符å<C2A6>ˆçº³å…¥æ ‡å‡†ã€?
dsConclusion: randomConclusion() // éš<C3A9>机ï¼?

// 特点�
- 1秒完�99�
- è¯<C3A8>æ<EFBFBD>®éƒ½æ˜¯"模拟è¯<EFBFBD>æ<EFBFBD>®"
- 判断结果éš<EFBFBD>机生æˆ<EFBFBD>

真实LLM（新�

// âœ?真实数æ<C2B0>®
dsPEvidence: "This study included adult patients with type 2 diabetes mellitus aged 18 years or older, which matches the population criteria."
dsReason: "The study population consists of T2DM patients, the intervention is an SGLT2 inhibitor (empagliflozin), the comparator is placebo, and the study design is a randomized controlled trial. All PICO criteria are met. The study reports on cardiovascular outcomes including MACE, heart failure hospitalization, and cardiovascular death, which are the outcomes of interest."
dsConclusion: "include" // AI真实判断�

// 特点�
- 33-66分钟完æˆ<EFBFBD>199ç¯?
- è¯<EFBFBD>æ<EFBFBD>®å¼•ç”¨æ–‡çŒ®åŽŸæ–‡
- 判断基于Prompt v1.0.0-MVP
- 准确率：60%（首次测试）

---

ðŸ”<EFBFBD> æ•°æ<C2B0>®éªŒè¯<C3A8>

验è¯<EFBFBD>方法

cd AIclinicalresearch/backend
node check-data.mjs

预期输出（真实数æ<EFBFBD>®ï¼‰

🔬 筛选结果样�
  [1] 文献: Assessment of Thrombectomy versus Combined...
      DeepSeek: include (P:match, I:partial, C:mismatch, S:match)
      Qwen:     exclude (P:mismatch, I:mismatch, C:partial, S:match)
      冲çª<C3A7>状æ€? conflict
      是å<C2AF>¦æœ‰è¯<C3A8>æ<EFBFBD>? DeepSeek=true, Qwen=true  âœ?

  è¯<C3A8>æ<EFBFBD>®ç¤ºä¾‹:
  - dsPEvidence: "The study population consists of..."
  - qwenPEvidence: "Patients with acute ischemic stroke..."

---

📊 è´¨é‡<C3A9>ä¿<C3A4>éšœ

已实现的质é‡<EFBFBD>措施

1. JSON Schema 验è¯<C3A8>:

   • 所有LLM输出必须通过Schema验è¯<EFBFBD>
   • ä¸<EFBFBD>å<EFBFBD>ˆæ ¼çš„输出会被拒ç»<EFBFBD>

2. 错误处ç<EFBFBD>†:

   • å<EFBFBD>•ç¯‡æ–‡çŒ®å¤±è´¥ä¸<EFBFBD>å½±å“<EFBFBD>整体任åŠ?
   • 详细错误日志记录

3. 进度追踪:

   • æ¯?0篇更新一次进åº?
   • 实时统计æˆ<EFBFBD>功/冲çª<C3A7>/失败æ•?

4. **å<>¯è¿½æº¯æ€?*:

   • 记录原始LLM输出（rawOutputï¼?
   • 记录Prompt版本（promptVersionï¼?
   • 记录处ç<EFBFBD>†æ—¶é—´ï¼ˆaiProcessedAtï¼?

---

🚀 测试步骤

Step 1: 准备测试数æ<C2B0>®

使用现有测试文件:
- PICOS: docs/.../测试案例的PICOSã€<C3A3>纳入标准ã€<C3A3>排除标å‡?txt
- Excel: docs/.../Test Cases.xlsx (199篇文�

Step 2: 执行测试

1. å<EFBFBD>¯åŠ¨å<EFBFBD>Žç«¯: cd backend && npm run dev
2. å<EFBFBD>¯åŠ¨å‰<EFBFBD>端: cd frontend-v2 && npm run dev
3. 访问: http://localhost:3001
4. 填写PICOS + 上传Excel
5. 点击"开始AIåˆ<C3A5>ç­›"
6. 等待30-60分钟�99篇�0秒）
7. 查看审核工作å<EFBFBD>?

Step 3: 验è¯<C3A8>结果

cd backend
node check-data.mjs

检查项:

• 所有文献都有筛选结æž?
• è¯<EFBFBD>æ<EFBFBD>®ä¸<EFBFBD>å†<EFBFBD>æ˜?模拟è¯<C3A8>æ<EFBFBD>®"
• è¯<EFBFBD>æ<EFBFBD>®åŒ…å<EFBFBD>«æ–‡çŒ®åŽŸæ–‡å¼•ç”¨
• 判断ç<EFBFBD>†ç”±è¯¦ç»†ä¸”符å<EFBFBD>ˆé€»è¾‘
• 冲çª<EFBFBD>检测准确（conclusionä¸<EFBFBD>å<EFBFBD>Œï¼?

---

âš ï¸<EFBFBD> 注æ„<C3A6>事项

API密钥é…<EFBFBD>ç½®

ç¡®ä¿<EFBFBD>环境å<EFBFBD>˜é‡<EFBFBD>å·²é…<EFBFBD>ç½?

# .env
DEEPSEEK_API_KEY=sk-xxxxx
QWEN_API_KEY=sk-xxxxx

APIé™<EFBFBD>æµ<EFBFBD>

• DeepSeek: 60 RPM（æ¯<C3A6>分钟请求数）
• Qwen: 60 RPM

当å‰<EFBFBD>ç­–ç•¥: 串行处ç<E2809E>†ï¼Œä¸<C3A4>会触å<C2A6>‘é™<C3A9>æµ?

æˆ<EFBFBD>本估算

• DeepSeek: ~$0.001/æ¬?× 199 = $0.20
• Qwen: ~$0.001/æ¬?× 199 = $0.20
• 总计: $0.40 / 次完整测è¯?

---

💡 优化建议

短期优化（Week 2 - Day 4-5�

1. 并å<EFBFBD>‘控制: 改为3个并å<C2B6>‘（33分钟 â†?11分钟ï¼?
2. 进度显示: å‰<C3A5>端轮询显示进度百分æ¯?
3. 错误é‡<EFBFBD>试: 失败的文献自动é‡<C3A9>è¯?æ¬?

中期优化（Week 3�

1. 消æ<EFBFBD>¯é˜Ÿåˆ—: 使用Bull Queue异步处ç<E2809E>†
2. 批é‡<EFBFBD>优化: 使用批é‡<C3A9>API接å<C2A5>£ï¼ˆå¦‚果有ï¼?
3. 缓存机制: 相å<C2B8>Œæ–‡çŒ®ä¸<C3A4>é‡<C3A9>å¤<C3A5>ç­›é€?

---

ðŸ“<EFBFBD> 相关文件

修改的文�

• backend/src/modules/asl/services/screeningService.ts â­?

ä¾<EFBFBD>赖的文件（已存在）

• backend/src/modules/asl/services/llmScreeningService.ts
• backend/src/modules/asl/schemas/screening.schema.ts
• backend/prompts/asl/screening/v1.0.0-mvp.txt
• backend/src/common/llm/adapters/LLMFactory.ts

测试文件

• backend/scripts/test-llm-screening.ts
• backend/scripts/test-samples/asl-test-literatures.json

---

🎉 æˆ<C3A6>果总结

已实�

âœ?真实LLM调用替æ<C2BF>¢Mockæ•°æ<C2B0>®
âœ?从项目读å<C2BB>–PICOS标准
âœ?å<>Œæ¨¡åž‹å¹¶è¡Œç­›é€? âœ?冲çª<C3A7>检测与标记
âœ?完整的日志追è¸? âœ?错误处ç<E2809E>†æœºåˆ¶

待优�

âš ï¸<EFBFBD> 处ç<E2809E>†æ—¶é—´è¾ƒé•¿ï¼?0-60分钟ï¼? âš ï¸<C3AF> 串行处ç<E2809E>†ï¼ˆå<CB86>¯æ”¹ä¸ºå¹¶å<C2B6>‘ï¼? âš ï¸<C3AF> å‰<C3A5>端进度显示（需优化轮询频率ï¼?

---

🔗 å<>‚考文æ¡?

• Prompt设计与测试完æˆ<EFBFBD>报告
• å<EFBFBD>’中数æ<EFBFBD>®æ³›åŒ–测试报告
• 任务分解

---

**报告�*: AI Assistant
日期: 2025-11-21
版本: v1.0.0