Autogen Notebooks

Python notebooks for playing around with autogen powered AI Agents.

Add API ENDPOINTS:

Create a file in same folder called OAI_CONFIG_LIST.json, with contents similar to:

[
  {
    "model": "gpt-4-1106-preview",
    "api_key": ""
  },
  {
    "model": "gpt-3.5-turbo",
    "api_key": ""
  },
  {
    "model": "local-model",
    "api_key": "NULL",
    "base_url": "http://localhost:1234/v1",
    "api_type": "open_ai"
  }
]

Note: The local-model is optional.

Notebook 101

1. Open the notebook file ipynb in VS Code
2. VS Code will suggest necessary extensions, install them.
3. Use "Run all" to set everything up
4. Tell it your task.
5. When re-running, it is recommended to delete .cache/ folder and change CACHE_SEED constant first.

Learnings

• system_message: Given to agents each time they are called on. Be specific about steps they should take. Be specific about the output you want.
• description: Used by the Manager of GroupChat to choose which agent should go next. it checks the last message, and is asked to pick a next agent, based on their descriptiond.
• retreive_agents: Needs both the agent and user_proxy to work. Only user_proxy agents can run functions. user_proxy agents don't have to interact with human.
• models: gpt3.5 is much faster but less reliable at orchestration. I've found that having Managers as gpt4 and then agents as 3.5 gives good results.