autogen/python/samples/agentchat_fastapi

AgentChat App with FastAPI

This sample demonstrates how to create a simple chat application using AgentChat and FastAPI.

You will be using the following features of AgentChat:

1. Agent:
   • AssistantAgent
   • UserProxyAgent with a custom websocket input function
2. Team: RoundRobinGroupChat
3. State persistence: save_state and load_state methods of both agent and team.

Setup

Install the required packages with OpenAI support:

pip install -U "autogen-ext[openai]" "fastapi" "uvicorn" "PyYAML"

To use models other than OpenAI, see the Models documentation.

Create a new file named model_config.yaml in the same directory as this README file to configure your model settings. See model_config_template.yaml for an example.

Chat with a single agent

To start the FastAPI server for single-agent chat, run:

python app_agent.py

Visit http://localhost:8001 in your browser to start chatting.

Chat with a team of agents

To start the FastAPI server for team chat, run:

python app_team.py

Visit http://localhost:8002 in your browser to start chatting.

The team also includes a UserProxyAgent agent with a custom websocket input function that allows the user to send messages to the team from the browser.

The team follows a round-robin strategy so each agent will take turns to respond. When it is the user's turn, the input box will be enabled. Once the user sends a message, the input box will be disabled and the agents will take turns to respond.

State persistence

The agents and team use the load_state and save_state methods to load and save their state from and to files on each turn. For the agent, the state is saved to and loaded from agent_state.json. For the team, the state is saved to and loaded from team_state.json. You can inspect the state files to see the state of the agents and team once you have chatted with them.

When the server restarts, the agents and team will load their state from the state files to maintain their state across restarts.

Additionally, the apps uses separate JSON files, agent_history.json and team_history.json, to store the conversation history for display in the browser.