[WillAdams]

Maybe?

But it's at my day job, and it's because I was able to write a prompt which automates having Copilot review uploaded scanned PDFs of invoices with checks (and the bank line obscured with a pen, so no PII) and then write a batch file which renames the files per a file-naming convention, removing the need to open them in batches of 50, find the Invoice ID, re-save using that filename, then quit and re-launch Adobe Acrobat (if left running, eventually I run into a bug where it stops saving files), then run a .bat file which renames based on Invoice ID as a filename.

Problem of course is I've been running into a limit of number of allowed files per 24 hr. period.

Even if it's not less work, it feels like less effort.

[pilgrim0]

Extracting structured data from PDFs is something people have been doing forever, way before LLMs, there are state of the art OCR tools out there you can use as a lib, it’s hardly something that justify using a LLM

[WillAdams]

The problem is, it's not very well structured.

The position of (for instance) the check # on a physical check varies quite widely, as can date position and format, and a fair number of them are still written by hand.

On top of that, I'm scanning thousands of these each year, and the invoice underneath the check ranges from the gamut of: "pristine copy just re-printed 'cause the customer didn't include one" through "bad inkjet photocopy of a photograph taken w/ a potato phone and then printed" and includes variations such as "customer included half-a-dozen invoices to be paid w/ one check, and if arranging them and the check _just so_ all will fit in the document camera window, saving a trip to the sheet feeder scanner"

A co-worker actually worked on this for a different program, one where actually sending in paperwork in good condition was expected and customary, but his system had a reject/failure rate of ~10% --- that would be almost 1,000 invoices each year for the program I am handling.

[fragmede]

Yes it is. It's a poorly defined problem, and it needs to be correct, and existing OCR tools aren't perfect. Are the end of the day, it comes down to money and time, and throwing them at ChatGPT is cheap and fast enough to be used.

[jorisw]

OCR is not the same as turning unstructured data (the freeform text OCRed from the PDF) into structured data.