使用说明
Supervisor Agent & MCP Tools – RAG System Backend
This repository contains the evolving backend for an enterprise-grade Retrieval-Augmented Generation (RAG) chatbot.
It uses a Supervisor Agent to route user queries to modular MCP-based tools, and is developed using a test-driven, tool-oriented agent architecture.
🗺️ Project Overview
The system is implemented in phases as outlined in the planning documents.
Key goals include:
- Modular orchestration of AI tools (retrievers, DB, web) via MCP
- LLM-agent routing logic with clear decision control
- Secure tool access (e.g., via On-Behalf-Of token flow)
- Full test coverage using TDD and clear stage separation
📁 Reference Documents
| File | Description | When to Use |
|---|---|---|
docs/project_background.md |
High-level vision, RAG architecture, staged roadmap | 📌 Use to understand the mission |
docs/supervisor_design.md |
Supervisor class architecture, config handling, CLI goals | 📌 Use during Stage 1 or agent implementation |
docs/supervisor_test_plan.md |
Full test cases per user story | 📌 Use during test writing or debugging |
docs/user_stories.md |
6 core user stories with acceptance criteria | 📌 Use when validating behavior coverage |
docs/openai_mcp_protocol.md |
In-depth guide to MCP, OpenAI function calling, security model | 📌 Use when designing real tools (Stage 2) or Supervisor-tool interfaces |
docs/Plan_for_Re-Authoring_GenAI_Components.md |
LangChain + Claude Code integration plan | 📌 Use when restructuring agents, tools, or orchestrators |
docs/Plan.md |
Overall development integration plan | 📌 Use to understand overall development steps and context of current tasks |
logs/phase*_*.md |
Claude Code outputs + test summaries | 📌 Use for traceability and phase review |
🧪 Development Flow
- Each stage is implemented using test-driven development.
- Stub tools are used early for fast testing; real MCP tools come later.
- Each phase’s results are logged in
logs/.
📌 Status (As of October 24, 2025)
| Phase | Status | Description |
|---|---|---|
| Phase 1 | ✅ Complete | Configuration module (config.py) – 10 tests passing |
| Phase 2 | ✅ Complete | Stub tool module (stubs.py) – 5 tests passing |
| Phase 3 | ✅ Complete | Router logic (router.py) – 8 tests passing |
| Phase 4 | 🟡 Next | Tool handlers (dispatch logic to stub tools) – 9 tests target |
| Phase 5 | ⏳ Pending | Supervisor agent (end-to-end query response logic) |
| Phase 6 | ⏳ Pending | Integration tests for full workflows |
| Phase 7 | ⏳ Pending | CLI interface (query entrypoint, config flags) |
Current Progress: 23/65 tests passing (35.4%)
🧠 Restarting Claude Code
If restarting the Claude Code terminal, use:
“Use the following reference files for this phase:
project_background.md,openai_mcp_protocol.md,supervisor_design.md,supervisor_test_plan.md.”
🔐 Secure Tool Usage
For tools requiring authenticated API calls (e.g., SharePoint, Azure DevOps), follow the OBO (On-Behalf-Of) pattern as defined in
docs/reference/Model_Context_Protocol_and_OpenAI_Chat_Completions_API.md.
📦 Structure (Stage 1)
supervisor/
├─ config.py
├─ router.py
├─ handlers.py
├─ agent.py
└─ tools/
├─ stubs.py
tests/
├─ test_config.py
├─ test_router.py
docs/
├─ *.md
logs/
├─ phase*_*.md
examples/
├─ mcp_server_example/
├─ langgraph_agent_example/
├─ openai_chat_example/
└─ mlflow_agent_model_example/
🧰 Real-World Examples Directory
A curated /examples directory provides working Python references that Claude Code and developers can use to understand real-world implementations of key technologies used in this project.
| Example | Source | Focus | Why It Matters |
|---|---|---|---|
| MCP Server Example | FastMCP | server.py shows @mcp.tool decorator and mcp.run() |
Demonstrates how to build an MCP-compliant microservice for tools (e.g., Document Retriever) |
| LangGraph Agent Example | LangGraph Example | graph_app.py defines and wires nodes |
Shows how to orchestrate multi-agent logic (used for Supervisor flow control) |
| OpenAI Chat Example | OpenAI Quickstart Python | app.py uses openai.ChatCompletion.create() |
Demonstrates ChatCompletion schema compatible with Supervisor’s chat layer |
| MLflow Agent Model Example | MLflow Official Examples | python_function.py wraps a Python class as a model |
Mirrors how we’ll expose agents as MLflow models in Databricks |
🔍 Usage Guidelines
- Use these examples as learning and grounding references for Claude when generating code.
- Each folder contains a
README.mdsummarizing key files and patterns to focus on. - These are not imported by the Supervisor — they exist to guide development and ensure architectural alignment.
LangChain + Supervisor Agent Integration
(Existing LangChain, LangGraph, and Supervisor content continues unchanged below…)
