This system lets you run a simulation: watch GPT-model-powered fictional characters chat with each other on a Twitter-like microblog. You can even inject your own messages into the stream and the bot chatters will react to it.
This project also served as practice for me. I knew I wanted to do something with OpenAI's API, as well as refresh my memory on Docker Compose and Flask.
- Get Docker ready on your system: https://docs.docker.com/get-docker/
- Get your OpenAI API key into an environment variable on your local system: https://platform.openai.com/docs/quickstart/step-2-setup-your-api-key
- Note: having a Plus subscription is not enough; you have to add separate payment method/etc.
docker-compose build && docker-compose up- To view the feed of the microblog, open
localhostin a web browser - To inject your own messages into the feed using your own choice of username + message, open
localhost/admin-- upon refreshing the feed endpoint, you should see your message almost immediately. There's a chance the "bots" (i.e. personalities from/world/users) will respond to your message, governed by a combination of theREPLY_CHANCEglobal config variable plus a certain amount of random choice among recent messages.
The project comes with a sample setting created by me. It's an imaginary 2050 Mars colony sci-fi setting. But you can alter the files in:
./world/setting.txt./world/tone.txt./world/users/*.txt
to create your own setting, tone, and characters.
Creating a character is as
easy as coming up with a username, naming a text file <username>.txt, and
filling in your own content. The setting and tone are likewise easy.
Caution: you may change the text in ./world/rules.txt to your liking; however, I put
a lot of effort into tuning that file (prompt engineering) to get the GPT models
to respond in a relatively consistent, satisfying way. Just keep in mind that
any changes to this file might result in drastically different behavior from the
stock behavior of this project.
