I’ve been facing a recurring issue with ChatGPT’s service that has been impacting my workflow significantly. Specifically, the following problems have become frequent:
Inaccurate Answers: The responses provided are often incorrect or misleading.
PDF Attachment Handling: ChatGPT seems to struggle with accurately reading and interpreting PDF attachments. It mixes up information or fails to extract data correctly.
Hallucinations: There are instances where the model appears to “make up” information, which is slowing down my processes.
These issues were not as prevalent before, but recently, they’ve been happening much more often. I rely on ChatGPT for efficiency, and this has become a roadblock for me and my team.
Here are my questions:
Is this a known issue, and are there any fixes or updates in progress to address it?
Do I need to delete and create a new account, or could there be a setting or configuration, way of use on my end that’s affecting performance?
Has anyone else experienced similar issues, and if so, how were they resolved?
I appreciate any guidance or assistance you can provide. I really value the service and would like to continue using it effectively.
I imagine you are just simply sending the PDF and asking the model to understand it, which it attempts to do using Code Interpreter. If the text is baked into the PDF or the pdf text extract mechanism has difficulties then you are asking a model to work with a bunch of noise.
By pre-processing the PDF and providing the model with a much more digestible format.
PDFs are meant to be printed. They can be incredibly hard to read and sometimes require an OCR. Which, ChatGPT does not offer this functionality.
Thank you for your feedback! I understand that PDFs can be tricky, but I want to emphasize that I’ve rarely needed to pre-process files in the past. ChatGPT handled them well for over a year without significant issues, even with no preparation.
The problems I’m encountering now seem unrelated to the file quality or structure, as the PDFs I use are uniform and standard for my industry. Since a few months ago, errors have become frequent, including:
Mixing up information between unrelated PDFs uploaded on the same day.
Failing to integrate corrections consistently within the same session.
These issues didn’t happen before, and my workflow hasn’t changed. Even when I correct ChatGPT, it often repeats the same errors or defaults to earlier misinterpretations.
I understand the limitations of PDF handling, but this behavior feels like a recent shift. Could there have been changes to how the model processes or retains session context?
Unfortunately ChatGPT is modified constantly without any notice. There is no known issue of it breaking with PDFs.
What usually happens is people setup very intricate instructions that they’ve iterated and tested over with the model until they reached a point of satisfaction.
Then, a new model comes along with potentially new system instructions, tooling, parameters, whatever, and it causes issues.
The solution is usually to revise your workflow and make it easier for the model to process the information you are providing to it.
I see what you mean. I’ve been experimenting with different approaches to avoid “confusing” ChatGPT. Just now, for example, I uploaded a clean PDF, asked for specific information, and received a response. Then, I started a clean chat screen, moved on to a completely new PDF, made it clear that it was a different document, and asked a question. However, ChatGPT still pulled information from the previous PDF in its response.
Even after I asked it to review and verify its answer, it insisted the information was correct, but it continued to reference data from the earlier PDF.
It’s been extremely frustrating. As you mentioned, I’ll continue revising my workflow to make it easier for the model to process the information I provide. I appreciate your insights and will keep refining my approach to improve the interaction with ChatGPT. Thank you again for your thoughtful input!
Are you using memories? Are you using a custom GPT? Are you uploading the document using the paperclip icon in the chat screen or uploading it directly to the CustomGPT?
There should absolutely be no left-over artifacts from your previous conversation when starting a new one with ChatGPT.