llm_agent_py

# LLM Agent High-performance LLM Agent system based on asyncio and MCP protocol, supporting multi-tool collaboration, RAG retrieval-augmented generation, and asynchronous task scheduling. ## Core Features ### 1. Agent Task Loop Scheduling Framework - **Asynchronous Task Scheduling**: Implemented a complete task loop scheduling mechanism based on asyncio - **Streaming Response Parsing**: Real-time parsing of LLM streaming output, supporting context management of tool call results - **Multi-Tool Collaboration**: Supports dynamic tool registration and automatic context injection ### 2. MCP Protocol Client - **Standardized Implementation**: Fully implemented Model Context Protocol client - **Tool Integration**: Successfully integrated 2 tool Servers (fetch, filesystem), including 15 native tools - **Lifecycle Management**: Supports dynamic registration, invocation, and resource cleanup of tools ### 3. Vector Database - **Implemented from Scratch**: Implemented a vector database based on pure Python, without relying on third-party vector libraries - **Cosine Similarity Algorithm**: Implemented an efficient semantic retrieval algorithm - **Performance**: - 10 documents: 1.25ms - 100 documents: 12.21ms - 1000 documents: 117.98ms - Linearly scalable performance ### 4. RAG Retrieval-Augmented Generation - **Embedding Pipeline**: Integrated BAAI/bge-large-zh-v1.5 pre-trained model - **Semantic Retrieval**: Top-K retrieval based on vector similarity - **Recall Accuracy**: Hit Rate@5 reaches 100% ## Core Module Description ### Agent.py Implements a complete Agent task loop scheduling framework: - Tool call routing and result management - Real-time parsing of streaming responses - Asynchronous task scheduling ### MCPClient.py MCP protocol client implementation: - Connection management with MCP Server - Tool list acquisition and invocation - Lifecycle management ### VectorStore.py Core implementation of the vector database: - Vector storage and indexing - Cosine similarity calculation - Top-K retrieval algorithm ### EmbeddingRetriver.py Embedding retriever: - Document embedding - Query embedding - Semantic retrieval ## Technical Highlights 1. **Implemented from Scratch**: Core components such as the vector database and MCP client are implemented from scratch, demonstrating underlying technical capabilities 2. **Asynchronous Architecture**: Fully asynchronous design, supporting high-concurrency scenarios 3. **Protocol Standardization**: Adopts the MCP protocol for standardized tool integration 4. **Performance Optimization**: Based on actual test data, performance is linearly scalable 5. **Engineering**: Complete test coverage, quantifiable performance data ## Future Optimization Directions - [ ] Use NumPy vectorization to optimize vector retrieval performance - [ ] Integrate professional vector databases (FAISS, Milvus) - [ ] Implement a caching mechanism to reduce redundant calculations - [ ] Support more MCP Tool Servers - [ ] Implement a multi-Agent collaboration mechanism