Custom GPT will not read full document in first try

My custom GPT is supposed to be an expert system of a technical book I wrote. The knowledgebase is a PDF, where all images have been removed, with 100.000 words plain text.
Whenever I ask something that is not a primary subject of teh text ( For example: Where can I get a glass funnel from?) GPT will answer that the document does not deal with sourcing equipment. Only if I insist GPT reads trough the whole document again and then find the section where this information can be obtained from. What can I do, to ensure that GPT really checks the whole document first before giving me a gnerl answer that this is a tech book and nothing else? I have tried various instructions (Read whole document first etc.) but nothing seems to work. Only if I insist (Try again…) GPT will return the correct answer.

1 Like

can you add a command something like before giving an answer to the user, make sure that the answer has come out of the PDF provided. If you don’t find that information in the PDF, you have generated that information. Instead, you should go back to the file and find the information relevant to the user, and provide the information only from the file. Again, please read the file whenever you need to provide an answer do not make up answers use the information in the PDF I have provided from you that is the only source you are to give to the user do not look up information on the Internet instead source it from my PDF

1 Like

@paulwlowder : Thank you for your feedback.
I now have found 2 modes that both do not satisfy me:
Mode 1: I do not explicitly instruct GPT to read the document fully before giving an answer. GPT returns a very general answer and does not know details about the text. Once I insist and retry, I will receive the correct very detailed answer quickly but it always requires asking a second time.
Mode 2: I instruct GPT to read through an index doc first, then the whole main document througly. It will take very long, because GPT rereads the main doc. Because of a time-out it only returns a vague information and refers to reading the book like:

While I was able to locate and verify that this information is covered in the document, I wasn’t able to reach the specific details within each section due to the time constraints in accessing the document. However, for comprehensive guidance on refining silver, I recommend referring to Chapter VIII in “(Book title)” for detailed instructions and methodologies.

I’m in the same configuration, and don’t find solution yet.
I’ve heard about a guy who creates a kind of setup, using vector table to avoid this.
I didn’t have the opportunity to test it.
I have a use case with a dozen of PDF so I need to find the right setup to improve this user experience.

Tell me if you find a way to fix this issue.

May be it should come by training the model :confused:

What kind of index do you use for your files? Is this a normal content page? And what did you do in the file?