B2B Document Extractor Rebuilt: Rule-Based vs. LLM Using...

Afterwards, we use pytesseract to read the image and extract the raw text via OCR (Optical Character Recognition). Put simply, OCR means that the tool "looks" at the image and tries to recognize letters from pixels. Quite similar to how humans decipher handwritten notes.

- In the second step, we use regex. These are regular expressions that search for specific patterns inside the text. For example, we can define: "Search for everything that comes after PO Number: ." Already in this second step, we can identify the first problem: What happens if the customer simply writes "Order Reference" instead of "PO Number: "?

In that case, the regex pattern finds nothing. What we can then do (or must do) is add a new rule. Execute Script 1 for Approach 1 Next, we create a new file called approach1_traditional.py with the following code that you can find in the GitHub-Gist inside the same folder: https://gist.github.com/Sari95/aa2be6938fbcb1c7f94b053d9046f55d Now we execute the file again inside the terminal: python approach1_traditional.py The Result of Approach 1 For Layout A, everything works perfectly: For Layout B?

Not a single field is recognized and all values return "None": And this is exactly where the problem lies. For every new customer, new regex rules would have to be written, tested, and deployed. With 200 customers, that means 200 different patterns.

And every time a customer slightly changes their form, the system breaks again. Approach 2: A New Way (pytesseract + Ollama + LLaMA 3) In this second approach, we keep the OCR step, but replace the rigid regex rules with an LLM: - pytesseract still reads the text from the PDF. - Instead of telling the code "Search for PO Number: ", we tell the LLM: "Here is an order document.