对话式聊天机器人

Conversational Chatbot

A persistent, multi-tool conversational chatbot built with LangGraph, Streamlit, and MCP.

https://github.com/user-attachments/assets/4177f1cf-70c2-41ff-bd76-73ce07428d14

Features

• Persistent chat history — conversations survive app restarts via SQLite checkpointing (per thread_id)
• Tool calling — LLM can invoke web search, stock prices, a calculator, and remote MCP tools
• Remote MCP integration — connects to a FastMCP expense tracker server with Bearer auth
• Multi-thread UI — sidebar lists and restores past conversations

Project Structure

chatbot/
├── chatbot_frontend.py        # Streamlit UI (chat interface + conversation sidebar)
├── chatbot_backend_sqlite.py  # LangGraph graph, tool loading, SQLite checkpointing
└── chatbot.db                 # SQLite database (auto-created; stores conversation state)

Quickstart

1. Install dependencies

pip install -r requirements.txt

2. Configure environment variables

Create a .env file in the project root:

# Required
GROQ_API_KEY=...           # LLM provider
MCP_API_KEY=...            # Remote MCP server (FastMCP) auth
ALPHAVANTAGE_API_KEY=...   # Stock price tool

# Optional
OPENAI_API_KEY=...
GOOGLE_API_KEY=...
HUGGINGFACEHUB_ACCESS_TOKEN=...

# LangSmith tracing (optional)
LANGSMITH_TRACING=true
LANGSMITH_ENDPOINT=https://api.smith.langchain.com
LANGSMITH_API_KEY=...
LANGSMITH_PROJECT=ChatBot-Project

3. Run the app

streamlit run chatbot/chatbot_frontend.py

How It Works

Conversation threads

Each chat session gets a unique thread_id (UUID). The sidebar lists all existing threads; clicking one restores its full message history via chatbot.get_state({"configurable": {"thread_id": ...}}). State is persisted using AsyncSqliteSaver backed by chatbot/chatbot.db.

LangGraph graph

The backend compiles a two-node graph:

• chat_node — calls the Groq LLM with tools bound via bind_tools
• tools (ToolNode) — executes any tool calls the model emits, then loops back to chat_node

Available tools

MCP tools are loaded via MultiServerMCPClient. If they fail to load (auth error, network issue), the app falls back gracefully to the remaining tools.

Architecture

flowchart TD
    U[User] -->|types message| S["Streamlit UI\nchatbot_frontend.py"]
    S -->|astream with thread_id| G["LangGraph App"]

    G --> CN["chat_node\nGroq LLM + bind_tools"]
    CN --> RT{"tool_calls?"}

    RT -->|No| END[Return assistant message]
    RT -->|Yes| TN["ToolNode"]

    TN --> T1[DuckDuckGo Search]
    TN --> T2[Stock Price / AlphaVantage]
    TN --> T3[Calculator]
    TN --> T4[MCP Tools / Expense Tracker]

    T4 -->|"Authorization: Bearer MCP_API_KEY"| MCP[(FastMCP Server)]
    TN -->|ToolMessage| CN

    CN & TN --> CP["AsyncSqliteSaver\nchatbot.db"]
    END --> S --> U

Troubleshooting

MCP tools not loading — verify MCP_API_KEY is set and https://efficient-purple-snipe.fastmcp.app/mcp is reachable.

Stock price errors — check that ALPHAVANTAGE_API_KEY is valid and not rate-limited.

Lost conversations — chat history lives in chatbot/chatbot.db. Deleting this file clears all threads.