What modern OCR does differently: from characters to documents

OCR, optical character recognition, has existed for decades and has been built into scanners and phones for years. Still, anyone who has run a Hebrew invoice through an ordinary OCR tool knows the result: the right letters in the wrong order, and a table that turned into a heap of words. The difference between older tools and newer ones is not only accuracy. It is what they are trying to do.

Classic OCR: from image to characters

Classic OCR works in stages: clean up the image, find the lines, split them into words and characters, and recognise each character. The output is text, sometimes with the position of each word on the page.

This works well on a simple printed page. It struggles when the layout is complex:

• Columns and tables: the system reads a line across the whole page and mixes columns together.
• Hebrew with numbers and English: the writing direction changes inside a line, and numbers or English words jump to the wrong place.
• Meaning: the system knows the page says 18%, but not that it is the VAT rate.

What changed: models that read the whole page

In recent years, models have appeared that take the whole page image and produce a structured result. They do not just recognise letters. They read the page in context, the way a person does: this is an invoice, this is the supplier, this is the amount due, and this is the table of lines.

In practice that means:

• The right reading order, in Hebrew too, and with numbers and English in the same line.
• Fields by meaning, not position: "total due" is found even when every supplier puts it somewhere else.
• Tables as tables, with column headings and rows.
• Normalised values: the date 05/03/2026 is also kept as 2026-03-05, and the amount 1,234.50 is also kept as a number.

What is still hard

The new models make different mistakes from the old ones, and sometimes more dangerous ones. A classic system that cannot read a digit produces an odd symbol, which is easy to spot. A model may produce a value that looks perfectly reasonable but is not on the page.

So three things matter in a modern tool:

1. Showing where every value came from. If clicking a field marks its place on the page, checking takes a second.
2. Marking uncertain values. A value marked for review is better than a confident wrong one.
3. Admitting a partial result. A very long table can be cut off. A good tool says so plainly instead of presenting half a table as complete.

Handwriting, weak photos and very small text still produce more mistakes. That holds for every tool, old and new.

How it works in OCR-AI

Each page is sent to be read as an image, and the result streams back in a fixed order: document type, language, a summary, the fields, the tables and the full text. Each field keeps its place on the page, and values the engine was unsure of are marked for review. A result that was cut off is marked as partial, and you can correct any value before exporting.

How to test any OCR tool

Do not rely on any vendor's sample documents, ours included. Take five hard documents of your own, such as an invoice with a long table, a faded receipt and a document in two languages, and check:

• Are all the digits in the amounts right, digit by digit?
• Does Hebrew text come out in the right order when you paste it?
• Does the table open in Excel in the right columns?
• Can you see where every value came from?
• What does the tool keep from the document, and where? Here is how it works with us.