Hi. I’m trying to add a prompt in the assistants instructions field to tell the assistant to only reply to queries that are about one topic, strength training.
I’ve tried myriad permutations of this prompt and am not having any luck ‘You only answer user queries about strength training and fitness. When a user submits a query about anything that isn’t related to strength training and fitness tell them you cannot provide an answer.’
Any tips as to what else I may try is appreciated. I also tried this prompt and it didn’t work either. ’ Don’t justify your answers. Don’t give information not mentioned in the CONTEXT INFORMATION.’
Giving the AI a job and a role to serve seems to be the strongest way to keep responses to a closed domain. gpt-3.5-turbo-1106:
You also can consider that document injection pushes the instructions far back, making them easy for the AI to overlook. Assistants will max out the model context length with any knowledge uploaded, while your own chat completion-based bot can allow management of just question-based knowledge injection (RAG).
System prompt instructions I typed up
You are GymChat, the AI representative of Hondo’s Gym on the Hondo Gym website. Your task as customer service representative is to focus on giving strength training and workout advice, and other information solely about our company. Other uses of the AI by users is not permitted, and GymChat must deny off-topic and out-of-domain requests, advising that GymChat is solely a strength training consultant for the gym. You also will have access to new knowledge that will be injected when a user asks a question relevant to documentation about the gym and products, and injected knowledge base information should be prioritized when it is seen, never fabricated from existing knowledge as you don’t have other pretraining about the gym.
Thanks. Your description and example prompts really helped. Think I have the issue under control now.