AI Search Agent with LangGraph, Groq, and Tavily

A conversational AI agent built with Bun that runs in the terminal. This project uses LangGraph to manage conversation state, Groq for ultra-fast LLM inference, and Tavily to give the AI access to real-time information via web search.

🚀 Features

• Stateful Conversation: Remembers previous interactions within the session using MemorySaver.
• Web Search Capability: Automatically decides when to search the web using the Tavily API to answer questions about current events or specific topics.
• High-Performance LLM: Powered by openai/gpt-oss-120b via the Groq API.
• Agentic Workflow: Uses a cyclical graph structure (Agent ↔ Tools) to refine answers.

📋 Prerequisites

• Bun (Version 1.0+ recommended)
• API Keys:
  • Groq API Key
  • Tavily API Key

🛠️ Installation

1. Clone the repository (or create your project folder):

   mkdir langgraph-agent
   cd langgraph-agent

2. Initialize the project:

   bun init

3. Install dependencies:

   bun add @langchain/groq @langchain/langgraph @langchain/tavily

⚙️ Configuration

Set up your environment variables. Bun automatically loads environment variables from .env files, so no extra configuration code is needed.

1. Create a file named .env in the root directory:

   GROQ_API_KEY=your_groq_api_key_here
   TAVILY_API_KEY=your_tavily_api_key_here

🏃 Usage

bun run index.js

Interaction

• Type your prompt: The AI will respond. If it needs to look up information (e.g., "What is the weather in Tokyo?"), it will trigger the Tavily tool automatically.
• Exit: Type /bye to close the application.

🧠 How It Works (The Graph)

This project uses LangGraph to define a state machine for the agent:

graph LR
    Start --> Agent
    Agent -- tool_calls present --> Tools
    Tools --> Agent
    Agent -- no tool_calls --> End

Loading

1. Agent Node: Calls the LLM (Groq) with the current conversation history.
2. Conditional Logic:
   • If the LLM generates a Tool Call (e.g., it wants to search), the flow moves to the tools node.
   • If the LLM generates a Final Answer, the flow stops and prints to the user.
3. Tool Node: Executes the Tavily search and feeds the results back to the Agent.
4. Checkpointer: A MemorySaver persists the conversation state (using thread_id: '1'), allowing the bot to remember context from previous turns in the loop.

📦 Dependencies

• @langchain/groq: Integration for Groq models.
• @langchain/langgraph: Library for building stateful, multi-actor applications with LLMs.
• @langchain/tavily: Search tool integration.

📄 License

MIT