React Agent from scratch

Overview

This project demonstrates a conversational agent workflow built using LangChain and LangGraph. The agent leverages an external tool to retrieve weather data from wttr.in and integrates a state graph to manage tool calls and agent responses in a loop until the conversation ends.

    +-----------+         
    | __start__ |         
    +-----------+         
          *               
          *               
          *               
      +-------+           
      | agent |           
      +-------+           
     *         .          
   **           ..        
  *               .       

+-------+ +---------+ | tools | | end | +-------+ +---------+

Key Features

• Agent Integration: Utilizes the ChatOpenAI model (gpt-3.5-turbo-1106) bound with custom tool functions.
• Weather Retrieval Tool: Provides a get_weather tool that fetches weather information based on a location input.
• State Graph Workflow: Implements a state graph using LangGraph's StateGraph to manage interactions between the agent and tool calls.
• Conditional Flow: The workflow conditionally routes responses—if the agent's message includes tool calls, it transitions to execute those tools, then resumes the agent conversation.

Project Structure

• Tool Definition:

  • get_weather(location: str): A tool function that calls the wttr.in API to fetch weather info. Returns a weather string if successful or an error message if not.

• Agent Setup:

  • The ChatOpenAI model from LangChain is instantiated with a specified model and temperature.
  • The model is then bound with the get_weather tool using the tool binding mechanism.

• Workflow Functions:

  • tool_node(state: AgentState): Processes tool calls from the latest message, invokes the corresponding tool(s), and wraps the result in a ToolMessage.
  • call_model(state: AgentState, config: RunnableConfig): Combines a system prompt with the conversation history and invokes the chat model, returning the model's response.
  • should_continue(state: AgentState): Examines the latest message to decide whether to route control to the tool node (if there are tool calls) or to end the workflow.

• Workflow Graph:

  • A state graph is built using StateGraph with nodes for the agent and tool functions.
  • The entry point is set to the agent node.
  • Conditional edges direct the flow: if there are tool calls, the graph transitions to the tools node; otherwise, it concludes the conversation.
  • After executing the tool, control flows back to the agent for further processing if needed.

How to Run

1. Setup Dependencies:
   Ensure you have the required libraries installed (e.g., LangChain, LangGraph, requests). You can install them via pip if necessary.

2. Execute the Script:
   Run the Python script. The script compiles the workflow graph, prints an ASCII representation of the state graph, and executes a sample query by sending an input message such as: