PolyBotConversation

PolyBotConversation is an experimental web application that allows you to have group chat conversations with humans and multiple LLM-powered chatbots with individualized "personalities."

For additional background, have a look at my blog post "PolyBotConversation – An Experiment with LLM Group Chats".

Important Notes

PolyBotConversation is purely experimental software and not designed for production. Most importantly, there are severe security and privacy implications. For example, conversations are not protected, and all users can see all conversations. Furthermore, due to the fact that "core memories" are, by default, generated for all conversations and bots, information will be retained after the deletion of chats.

Key Features

• Multi-user (group) chats with both humans and LLM-powered chatbots with individual "personalities"
• @mention and conversation-based triggers for chatbot responses
• Automatically establishing so-called "core memories" that shape a bot's "personality" across conversation
• Automatically naming conversations (LLM-powered)

Setup

Python Environment

• Create a Python environment (e.g., using venv, conda, poetry or uv).
• Install the dependencies in requiements.txt.

OpenAI Key

You will need to set your OpenAI Key (if you are using OpenAI) as an environmental variable.

Windows

set OPENAI_API_KEY=sk-...

Linux

export OPENAI_API_KEY=sk...

Configuration and Start (Development Environment)

First, have a look at both the settings.py and the prompt_templates.py. Here, you can make changes to how the system functions.

Then, run the setup script:

python manage.py setup

Now you can run the development server:

python manage.py runserver

In a different terminal, run the following in order to start the task manager:

python manage.py qcluster

Development

Feel free to tinker around with this and make suggestions! PRs are absolutely welcome!

Notes for Developers

Here are a few notes about the design of the application that hopefully will help you to get started!

• The application is based on Django and uses htmx for the chat itself. Bootstrap is used as a CSS framework. Both htmx and bootstrap are shipped here in the static/js folder. Of course, they don't fall under the same license as this project.
• The application makes heavy use of django-q2. tasks.py holds the key tasks that are being performed by the task manager. Some tasks are being issued on the fly, e.g., the generation of core memories in views.py.
• There is a simple "prompt library" in prompt_templates.py. These are populated using .format() before usage with the API.
• bot.py contains functions for generating messages and replying in conversations. llm.py contains functions for interacting with the OpenAI (compatible) endpoint and other LLM tasks.
• The general flow looks like this: A task (tasks.py) runs triggers from triggers.py. These use functions from bot.py (which uses llm.py) in order to generate new message.

Linting and Formatting

isort and ruff are used for linting and formatting. I don't particularly like PEP-8's maximum line length, so ...

isort .
ruff check
ruff format --line-length 200