Errors in reading knowledge base in the custom GPT

Hi! I am trying to load knowledge base as PDF files in my customized GPT, but I keep getting errors when the chatbot tries to access it during the testing phase. It takes very long to read it and then it gives me an error. Could it be that the PDF is too long? is there a limit on the number of words?


How big is the PDF? can you post a screen shot of the error?

Sure. Here it is:

It is very strange. Because I tried to ask the same question some minutes later and it managed to read from the knowledge base and provide me with an answer. It seems very erratic…

The PDF is not particularly big in terms of size (around 6 MB) but it is very packed with words and written in two columns per page (it is the civil code for the Italian law). I broke it down into sections and told the GPT to retrieve the right section depending on the topic we are discussing so that it does not have to parse a huge amount of words, in case the 25000 words limit applies.

What about a different PDF? just for testing?

1 Like

Could be a solution. However, it is not very easy to find PDF versions of the civil code (especially because there is an updated one every year and I would like to use the latest). I will try to look at some alternative versions…

Can you copy the contents from your PDF and upload it as a .txt file instead?
That should make it more linear as well and easier for the GPT to understand

1 Like

I mean ANY other PDF, just to see if you get the same error.

1 Like

That could be a good solution, thanks! I will try it out!

1 Like

I see. I will try to do it with an older version and see what it does. Thanks for the tip.

I followed your tip and converted the PDFs into text and redefined the two columns per page into one. It seems that it is working better now. So far I got no problems in reading the document. I get this network error, though

Might it be a problem on the server side?

1 Like

Possibly, I’ve been randomly getting those too when in the playground and the GPT Plus interface. The Assistants API has been a bit more reliable for my testing, but the GPTs work okay too.

Glad to hear converting it to .txt helped a bit though!

1 Like

Yes, Network Errors are usually the result of an overly busy server, the entire system is under very heavy load after Devday.


Thank you very much to both! I hope no more parsing problems show up. In case I will knock on this door again :-).

Reasonably so, what they announced was crazy! Watching live was a blast.


I got to sit 10 feet from Sam! The atmosphere was awesome.


I am stupidly jealous. Are you part of the actual OpenAI team or an extremely lucky/skilled API user?

I live and work in the UK now but I’ve done a bit of a deep-dive over the past few months, I would absolutely love to get involved with the team behind it. There’s a ton of little things that would make these tools insanely valuable to the collision repair industry, let alone the automotive sector as a whole.

I had a working GPT this morning using all of my knowledge base.

An hour later, I tested it with a friend and it stopped using any of the knowledge base and was giving generic answers. I’ve removed all the documentation and added it back and have changed file formats and still not working after 12 hours.

I’m not sure what to do from here.