AutoGen Labs

Useful components and experiments to help AutoGen devs.

TL;DR

To run the sample app AutoGenGuiChat, do the following (ideally, in your Python virtual environment):

1. Install requirements:

   git clone https://github.com/radman-x/AutoGenLabs.git
   cd AutoGenLabs
   pip install -r requirements.txt
   

2. Copy OAI_CONFIG_LIST_sample to OAI_CONFIG_LIST and add your OpenAI/Azure key(s). Read, then remove comments in that file.
3. Run:

   panel serve AutoGenGuiChat.py
   

4. Point browser at http://localhost:5006/AutoGenGuiChat
5. Enjoy chatting with AutoGen!

Components

• AutoGenGuiChat -- sample application for using AutoGenChatView as graphical user interface for AutoGen
• AutoGenChatView -- re-usable chat view component

Features

• "ChatGPT-like" user experience
• Multi-turn, interactive AutoGen graphical user interface (GUI) via AutoGen async.
• Pure Python -- no HTML/Javascript/React/backend-frontend, etc.
• Based on Panel
• Lightweight, easily customizable and hackable for embedding in your own applications
• Displays usage stats with total cost and tokens in console (in DEBUG mode)

Test Prompts

These prompts are know to work in AutoGenGuiChat and can be run in the same session:

• "tell a joke"
• "write a very short poem"
• "write a python function to list current directory"
• "tell another joke"
• "write a very short story"
• "summarize conversation so far"

Customizing

There are various methods to customizing these components including:

• Override AutoGenGuiChat app class and write your own build_autogen_flow() method. This lets you create more advanced AutoGen flow e.g., like adding an engineer, critic, etc..
• Hacking the AutoGenGuiChat app class directly, e.g., to customize the build_autogen_flow() method without overriding the class.
• Hacking the AutoGenChatView component class directly.
• In the build_autogen_flow() method, you can modify agent system messages, descriptions (ie, used for GroupChat next speaker selection) and avatars. Avatars can use any emoji unicode character such as:
  • Smileys & People : https://emojipedia.org/people/
  • Animals & Nature : https://emojipedia.org/nature/
  • Foor & Drink : https://emojipedia.org/food-drink/
  • Activity : https://emojipedia.org/activity/
  • Travel & Places : https://emojipedia.org/travel-places/
  • Objects : https://emojipedia.org/objects/
  • Symbols : https://emojipedia.org/symbols/
  • Flags : https://emojipedia.org/flags/

Debugging

• To run/debug in VSCode, create a panel launcher configuration in VSCode's launch.json, eg:

      {
          "name": "panel serve",
          "type": "python",
          "request": "launch",
          "program": "-m",
          "args": [
              "panel",
              "serve",
              "${relativeFile}",
              // "--show" // This will open in Edge (ie, default OS browser)
          ],
          "console": "integratedTerminal",
          "justMyCode": false
      },
  

• Then open and select AutoGenGuiChat.py tab in VSCode and hit F5.

Know issues

• Function calls are not displayed in view
• This produces a warning message like "GroupChat is underpopulated with 2 agents...". This can be safely ignored.
• Occasionally, the "summarize..." test prompt above is ignored; usually after repeating it again, it works.
• Occasionally, the interaction between the agents is wrong (eg. message sent to Assistant instead of Admin) so no further human input can occur. Get "There is currently no input being awaited." in console.
• Numerous standard icons in Panel ChatInterface are not yet implemented: "liking" messages, clear history, etc.
• Add better documentation