I’ve been playing with the ChatGPT Turbo API in the Playground, and I’ve experimented by trying to restrict dialogue to a particular topic. I’m guessing that this isn’t an unusual thing to try to do, after all, integrating ChatGPT into a corporate chatbot would require the answers given to be reliable and professional.
I’m getting a lot of hallucinations. When the AI doesn’t know something, it just makes it up. It also wanders off topic. I’ve tried prompting in the system context and then reinforcing in the first message of the user context in this kind of fashion:
SYSTEM: You are a chatbot for Acme Corporation. You will only answer questions about Acme Corporation.
USER: What follows is information about Acme Corporation. When you are asked about Acme Corporation use the information below, and only the information below to formulate a response. If you are asked to discuss a topic that does not relate to Acme Corporation say “I don’t know about that”. If you are asked a question that is not answered in the information below, direct the user to the Acme Corporate website.
Address: 123 Letsby Avenue, London
Acme is a blah blah blah…
Acme produces blah blah blah…
Acme TOS is blah blah blah…
Am I being unrealistic in my expectations, or do I just need to learn more about how to prompt the AI?