LangGraph Basics

A minimal LangGraph example that powers an RPG-style chatbot. Each user turn is:

1. summarized into a running campaign recap,
2. classified into a message type (exploration, combat, or social interaction),
3. routed to a specialized agent that responds in that mode.

Features

• Structured classification with Pydantic (MessageClassifier)
• Persistent state across turns (messages, type, summary)
• Conditional routing to specialized agents
• Concise session summary updated every turn

Requirements

• Python 3.13+
• A Google Gemini API key (set in .env as GEMINI_API_KEY)

Setup

# from the repo root
uv sync

Create a .env file in the project root:

GEMINI_API_KEY=YOUR_KEY_HERE

Run

python src/main.py

Type messages at the prompt. Use exit to quit.

How it works (high level)

• summarize_turn maintains a compact session summary
• classify_message uses structured output to pick a message type
• router sends the flow to one of three agents
• Each agent responds with domain-specific guidance

Project layout

.
├── src/
│   └── main.py
├── pyproject.toml
└── README.md

Notes

• The model is set to gemini-2.5-flash in src/main.py. Adjust if needed.
• The summary is printed each turn; comment it out if noisy.