To output results using only the contents of text files uploaded to GPTs

Multiple-choice problems are created with content limited to the content of text files uploaded to myGpts, except for the basic knowledge of the web and GPT.
However, sometimes GPT is generated by quoting the basic knowledge of the web or gpt, that is, what is outside the text.
Any tips to prevent this?

I’m having the same problem, any suggestions on how to solve it by limiting the answers to only the contents of the training file?