RAG MCP agentAPI

设置和运行说明-实验7

快速参考

开始一切

docker run -d -p 6379:6379 --name devops-redis redis:7-alpine
python mcp_server.py  # Terminal 1
python agent_api.py   # Terminal 2

测试一切

curl http://localhost:8000/health
curl -X POST "http://localhost:8000/rag/ingest?directory=data/docs"
python test_mcp_tools.py
k6 run k6/load_test.js

停止一切

# Press Ctrl+C in both terminals
docker stop devops-redis

先决条件

pip install fastapi uvicorn chromadb pydantic groq python-dotenv langgraph langchain-core redis requests

附加要求：

• Docker（适用于Redis）
• k6（用于负载测试）

______________________________________________________________________

步骤0：启动Redis

docker run -d -p 6379:6379 --name devops-redis redis:7-alpine

验证Redis：

docker exec -it devops-redis redis-cli ping
# Expected: PONG

______________________________________________________________________

步骤1：配置环境

创建 .env 文件：

GROQ_API_KEY=gsk_your_key_here
REDIS_URL=redis://localhost:6379
BASIC_RPM=10
PRO_RPM=60
VIP_RPM=300

______________________________________________________________________

步骤2：启动MCP服务器（终端1）

python mcp_server.py

预期产量：

✓ Database initialized
✓ Sample server data seeded
✓ MCP Server started

______________________________________________________________________

步骤3：启动代理API（终端2）

python agent_api.py

预期产量：

✓ Connected to Redis at redis://localhost:6379
✓ Groq LLM configured
✓ LangGraph workflow initialized
✓ MCP client connected to server
INFO:     Uvicorn running on http://0.0.0.0:8000

______________________________________________________________________

第4步：摄取文件

curl -X POST "http://localhost:8000/rag/ingest?directory=data/docs"

预期产量：

{
  "status": "success",
  "documents_ingested": 45,
  "message": "Ingested 45 chunks from data/docs"
}

______________________________________________________________________

步骤5：测试系统

健康检查

curl http://localhost:8000/health

预期： 所有值 true

测试查询（RAG）

curl -X POST "http://localhost:8000/agent/query" `
  -H "Content-Type: application/json" `
  -H "X-Client-ID: basic-test" `
  -d '{\"message\": \"What is the deployment procedure?\"}'

预期： 带有引用和速率限制标头的HTTP 200

测试速率限制

# Send 15 requests (over 10 RPM limit)
for ($i=1; $i -le 15; $i++) {
  Write-Host "Request $i"
  curl -X POST "http://localhost:8000/agent/query" `
    -H "Content-Type: application/json" `
    -H "X-Client-ID: basic-ratelimit-test" `
    -d '{\"message\": \"Test\"}'
  Start-Sleep -Milliseconds 500
}

预期： 前~10个获取HTTP 200，其余获取HTTP 429

测试MCP工具

# Create ticket
curl -X POST "http://localhost:8000/agent/query" `
  -H "X-Client-ID: vip-test" `
  -H "Content-Type: application/json" `
  -d '{\"message\": \"Create a critical ticket for database outage\"}'

# Get ticket
curl -X POST "http://localhost:8000/agent/query" `
  -H "X-Client-ID: vip-test" `
  -H "Content-Type: application/json" `
  -d '{\"message\": \"Get ticket 1\"}'

# Append note
curl -X POST "http://localhost:8000/agent/query" `
  -H "X-Client-ID: vip-test" `
  -H "Content-Type: application/json" `
  -d '{\"message\": \"Append note to deploy-2024-01: Success\"}'

______________________________________________________________________

步骤6：运行k6负载测试

安装k6

choco install k6

运行测试

# Basic tier test
k6 run k6/simple_test_basic.js

# Comprehensive test
k6 run k6/load_test.js

# Mixed clients test
k6 run k6/simple_test_mixed.js

预期： 超过限制时，请参阅429条回复

______________________________________________________________________

步骤7：运行自动测试

测试MCP工具

python test_mcp_tools.py

测试RAG精度

python eval_rag.py

______________________________________________________________________

建筑

Client (k6/curl) → Rate Limiter (Redis) → LangGraph Orchestrator
                                                ↓
                                    ┌───────────┴────────────┐
                                    ↓                        ↓
                                RAG System              MCP Tools
                                    ↓                        ↓
                                ChromaDB                SQLite DB

具有速率限制的请求流

1. Client → API: Request with X-Client-ID header
2. Rate Limiter → Redis: Check bucket state (Leaky Bucket)
3. If allowed → Orchestrator: Process request
4. Orchestrator → RAG/MCP: Based on intent
5. Response → Client: With rate limit headers

______________________________________________________________________

技术栈

核心组件

• API：带限速中间件的Python FastAPI
• 速率限制器：Redis+漏桶算法
• 矢量数据库：ChromaDB（嵌入式）
• LLM：Groq（免费套餐）
• MCP服务器：Python（stdio传输）
• 编排器：LangGraph（基于图形的工作流）

测试

• 负载测试图形 k6
• 单元测试：Python单元测试
• 集成测试：自定义测试套件

______________________________________________________________________

项目结构

lab7-devops-agent/
├── agent_api.py              # FastAPI server with rate limiting
├── rate_limiter.py           # Redis + Leaky Bucket implementation
├── devops_orchestrator.py    # LangGraph orchestrator
├── rag_system.py             # RAG with ChromaDB
├── mcp_server.py             # MCP server (3 tools)
├── mcp_client.py             # MCP client
├── test_mcp_tools.py         # Automated tool tests
├── eval_rag.py               # RAG evaluation script
├── docker-compose.yml        # Redis + API services
├── Dockerfile                # API container image
├── requirements.txt          # Python dependencies
├── .env                      # Environment variables
├── k6/
│   ├── load_test.js         # Comprehensive k6 test
│   ├── simple_test_basic.js # Basic tier test
│   ├── simple_test_pro.js   # Pro tier test
│   └── simple_test_mixed.js # Mixed clients test
├── data/
│   └── docs/                # DevOps documentation (10+ files)
│       ├── deployment_procedure.md
│       ├── incident_management.md
│       ├── server_maintenance.md
│       ├── monitoring.md
│       ├── backup_policy.md
│       └── ... (5+ more)
└── tests/
    ├── test_queries.json    # RAG test queries
    └── eval_results.json    # Evaluation results

______________________________________________________________________

费率限制级别

客户端ID示例：

• basic-user-123
• pro-company-456
• vip-enterprise-789

______________________________________________________________________

MCP工具（3个工具）

1. 创建_票 -创建事件/问题工单

- 输入：摘要、详细信息、优先级 - 输出：ticket_id，状态

1. 获取门票 -按ID检索门票详细信息

- 输入：ticket_id - 输出：票详细信息或未找到

1. 附录_注释 -为实体添加注释

- 输入：entity_id，备注 - 输出：note_id，确认

______________________________________________________________________

Docker编写替代方案

使用Docker Compose而不是手动步骤：

# Start all services
docker-compose up -d

# Check status
docker-compose ps

# View logs
docker-compose logs -f agent-api

# Ingest documents
curl -X POST "http://localhost:8000/rag/ingest?directory=data/docs"

# Stop services
docker-compose down

______________________________________________________________________

故障排除

Redis未连接

docker ps | findstr redis
docker start devops-redis

速率限制不起作用

curl http://localhost:8000/health
# Check "redis_ready": true

MCP服务器没有响应

# Restart MCP server (Terminal 1)
# Press Ctrl+C, then:
python mcp_server.py

重置所有速率限制

docker exec -it devops-redis redis-cli FLUSHALL

端口8000已在使用中

netstat -ano | findstr :8000
taskkill /PID 
 /F

______________________________________________________________________

成功检查表

• ✅ Redis正在运行： docker ps | findstr redis
• ✅ 健康检查：全部 true 在 /health
• ✅ 摄入的文件：40+块
• ✅ RAG工作：查询返回引用
• ✅ MCP工具工作：所有3个工具都经过测试
• ✅ 速率限制：基本层在10个请求后获得429
• ✅ k6测试：负载测试显示预期的速率限制
• ✅ 不同级别：专业/VIP有更高的限制

______________________________________________________________________