Improve PDF OCR performance from ChatGPT

[tb]

Hi there,

I’m trying to setup an automation to acquire invoices from a mailbox, and summarize them using ChatGPT.

My issue is that ChatGPT struggles to extract accurate information from the invoices. The invoices are in clean one-page PDFs. Somehow, ChatGPT invents some information and fails to acquire most information.

Is there a known fix for this? I know that other specialized third-party invoice analysis tools exist, but if possible I’d prefer to stick to ChatGPT.

Automation overview below (two branches to deal with PDF and image files separately).

Thank you so much in advance!

[Milan_Vasarhelyi]

Which model do you use? Do you have the same issue with chatgpt.com or just the API?

[Stoyan_Vatov]

Hey there,

please have in mind that ChatGPT is NOT a PDF OCR platform. You can try creating an agent and training it with your own data to improve its performance or switch to a different tool actually designed to do PDF OCR.

[samliew]

Welcome to the Make community!

Try using the Dumpling AI “Extract data from PDF with AI” module:

Extract structured data from one or more PDF files with multimodal AI (10+ credits)

For more information on how to set this up, see

• Extract text content of pdf - #2 by samliew
• How much do AI tools cost? - #3 by samliew

Hope this helps! Let me know if there are any further questions or issues.

— @samliew

P.S.: Investing some effort into the Make Academy will save you lots of time and frustration using Make.

[Andrei_Unterzeger]

Can you tell me what role the router plays in your scenario?

[Andrei_Unterzeger]

I managed to complete this task in this way:

At the output, the data is loaded into a Google table.
Filter for PDF attachments.

There were problems with the Dumpling AI settings. I did some magic and watched the developer’s video tutorials. The screenshot shows the settings.