LangGraph Agent Exercises

LangGraph Agent Exercises is a hands-on playground for mastering stateful AI agents using the LangGraph framework in Python. Through sequential examples—from a simple command-line bot to sophisticated RAG-powered assistants—you’ll learn to design resilient conversation flows, bind custom tools, and ground language models with real data.

This repository covers essential AI engineering concepts:

• Structured state management with AiMessages, HumanMessages, ToolsMessages, SystemMessage and dynamic workflows via LangGraph's StateGraph.
• Modular tool integration, custom message routing, and prompt engineering best practices.
• Retrieval-Augmented Generation (RAG) to enhance accuracy by combining retrieval and generation.

By completing these exercises, you’ll gain practical experience in building scalable, maintainable LLM applications and understand how modern AI agents orchestrate retrieval, reasoning, and generation to deliver context-aware, factual responses.

Key Concepts

LangChain & LangGraph
LangChain lets you wire together prompts, embeddings, and tools to build LLM-powered apps. LangGraph adds a state-graph layer, so your agents can manage conversation flows, tool calls, and data pipelines as clear, visual workflows.

Retrieval-Augmented Generation (RAG)
RAG supercharges your AI’s knowledge by combining retrieval of real documents with the power of generative models. Imagine your agent has a library of PDFs and articles:

1. Query Understanding: When you ask a question, the agent turns your words into a numerical vector (embedding).

2. Smart Search: It then “zooms in” on the most relevant passages in its document store, pulling out the top 3–5 snippets.

3. Context Assembly: Those snippets get stitched into your prompt, so the AI has actual facts right in front of it.

4. Knowledgeable Answer: Finally, the LLM writes a response informed by those precise excerpts.

RAG gives you both accuracy (factual grounding) and creativity (natural-language answers), making it perfect for anything from customer support to research assistants.

OpenAI ChatGPT API Setup & Usage

1. Install SDKs:

   pip install openai langchain langchain_core langchain_openai

2. API Key:

   • Create an .env file at project root with:

     OPENAI_API_KEY=sk-...

   • Load keys in Python:

     from dotenv import load_dotenv
     load_dotenv()

3. Instantiate Client:

   from langchain_openai import ChatOpenAI
   
   client = ChatOpenAI(
       model='gpt-4.1-nano',
       temperature=0.0,
   )

4. Send Messages:

   from langchain_core.messages import SystemMessage, HumanMessage
   
   msgs = [SystemMessage(content='You are an expert helper'), HumanMessage(content='Hello!')]
   reply = client.invoke(input=msgs)

Repository Structure

• AgentExercise.ipynb – Jupyter notebook with all five agent exercises
• requirements.txt – Python dependencies
• .vscode/ – VS Code settings and launch configurations

Prerequisites

• Python 3.10+ (tested on 3.13.2)
• Virtual environment tool (venv, conda, etc.)

Setup

# Create & activate a virtual environment
python -m venv .venv
# macOS/Linux
source .venv/bin/activate
# Windows
.venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

Running the Exercises

1. Open AgentExercise.ipynb in VS Code or Jupyter.
2. Execute cells in order.
3. View graph diagrams inline via display(Image(app.get_graph().draw_mermaid_png())).

Exercise Overview

Agent 1: Personalized Compliment

• State: { name: str }
• Node: compliment_node
• Output: "Hi {name}, you're doing an amazing job learning LangGraph!"

Agent 2: List Processor

• State: { name: str, values: list[int], operation: "+"|"*" }
• Node: process_values
• Performs addition or multiplication on values.

Agent 3: User Profile Graph

• State: { name: str, age: int, skills: list[str], result: str }
• Three nodes: greeting → age description → skill listing.

Agent 4: Dual Operation Graph

• State: { number1, number2, number3, number4: int; operation1, operation2: "+"|"-"; result: str }
• Two conditional routers apply + or - twice and combine results.

Agent 5: Auto Guessing Game

• State: { player_name, guesses: list[int], attempts, lower_bound, upper_bound, target_number, hints }
• Looping graph: guess → hint → continue/exit until correct or max 7 attempts.

Customization & Extension

• You can extract each graph into standalone Python modules under src/ and add unit tests via pytest or unittest.
• Extend state schemas by defining new TypedDict entries.

Module Details

simple_bot.py

A foundational example (Agent I) showing how to integrate LangChain’s ChatOpenAI model with a graph-based workflow. It:

• Defines an AgentState using a list of HumanMessage objects.
• Initializes a GPT-4.1 nano model for language generation.
• Builds and compiles a StateGraph to manage message processing.
• Demonstrates modular design and command-line interaction.

chat_bot.py

Agent II: a full-featured chatbot with conversational memory. It:

• Uses TypedDict for structured state (HumanMessage/AIMessage).
• Maintains entire history and routes messages via StateGraph.
• Loads API credentials from .env and logs conversation to a file.
• Provides an interactive loop for context-aware AI responses.

react_agent.py

Agent III: a ReAct (Reasoning + Acting) agent. It:

• Shows how to create and bind custom tools (@tool functions).
• Constructs a ReAct graph to interleave reasoning and external tool calls.
• Routes ToolMessage and SystemMessage through StateGraph.
• Handles dynamic message updates and error-resilient tool integration.

brainstormer.py

Agent IV: Brainstormer agent for idea generation. It:

• Generates creative prompts and appends ideas to the session state.
• Persists brainstorming output to an external file.
• Integrates with LangChain tools and a graph-based session flow.
• Illustrates functional patterns in state management.

rag_agent.py

Agent V: Retrieval-Augmented Generation (RAG) agent. It:

• Loads documents (e.g., PDF) and splits text for indexing.
• Uses OpenAIEmbeddings and a Chroma vector store for similarity search.
• Enhances model context with retrieved content before generation.
• Manages retrieval + generation loops via StateGraph.

Contributing

Please follow our coding style guidelines and PR description template.

License

Licensed under the terms of the MIT License.