Providing context to the Chat API before a conversation

[shrewd quest]

I’m using the API playground right now to test providing user demographic information to GPT. After reading the Chat Guide I figured the best way to do this with 3.5 is to use a User message. So wrote up a few sentences for the initial user message:

“Use second grade English. I am learning second grade math. Never provide the answer but help me walk through the problem. Use the english language, but keep in mind that my native tongue is Spanish. When I give an answer double check that it is correct given the original question, even if my explanation is correct.”

The results are ok but I’m trying to figure out the best way to give GPT this kind of info in order to reduce it’s mistakes. Using multiple user messages seems to make the assistant more likely to ignore the provided context and less empathetic.

Does 1st person or 3rd person language make a difference? Temperature also seemed to make the assistant more likely to ignore the initial context, has anyone else found this to be true?

[bold perch]

Try providing the context in one user message separated by new line characters rather than in new user messages, that way you will only have two user messages - one for context and one for the query.

Alternatively, try appending the user-provided query onto a preset user message.

For instance:

query = "Do this but don't do that. Name: Jan /nAge:23/nThe user has asked the following: " + user_input

You could also try messing with the system message to instruct it there.

[shrewd quest]

The appending strategy seems worth a try! Now that you've mentioned it I've noticed good results doing that when using chatGPT.