Correct retrieval of figures from uploaded files

I created a Custom GPT designed to advise people on the characteristics and features of specific products. For this purpose, I uploaded several PDFs containing Instructions For Use for some products. However, when I ask this chatbot about certain characteristics, it provides incorrect values and even different answers to the same questions.
Here is an example of a table from one of the PDFs:

Has anyone tried something similar?
Is there a better way to present this information during fine-tuning?
I have spent two days attempting to prompt and fine-tune, but the issue remains unresolved.
Would using a different file format, like JSON, help the model more accurately retrieve product characteristics?"

I removed all knowledge and created GPTs related to the subject. Then send the document via chat.

At first, I thought of doing online knowledge, but felt that there would probably be no problems in releasing the website.


new information that might be helpful on this issue. I used Text Extractor - GPT from Open AI and specify for it which data to extract from my pdf with Instruction For Use. So for some prodcuts it extracted the wrong data. Now I am confused becasue I thought that my wrong results were because system were sort of halucinating, but when extractor (which must be just extract the data from the tables) also mixed up data, I do not know what to think… Maybe some internal (not visible to me) markups of tables in my pdf somehow wrong…

1 Like