Vision models got good at reading documents, and the result is that a lot of people stopped using OCR. For a whole class of documents that was a mistake.
What a vision model is good at
Layout. A vision model sees the page, so it understands that a figure caption belongs to the figure and that a table has headers. That is a genuine advantage over OCR, which produces a stream of characters and no structure.
Reading a table. Where the structure is visual and irregular, a vision model often gets it right where a text-based converter produces a mangled pipe table.
A page that is mostly a picture. A scanned form, a chart with text in it, a page with a diagram and a paragraph. OCR gives you the paragraph and nothing else.
What OCR is still better at
Speed and cost. A page through OCR is a fraction of the cost and a fraction of the time of a page through a vision model, and for a corpus of 50,000 pages that difference is the whole project.
Exactness. OCR gives you the characters. A vision model gives you what it read, which is usually right and occasionally confidently different. For a serial number, a figure in a table, a version string, a legal citation — you want the characters.
Auditability. You can point at the word an OCR engine was unsure about. A vision model does not have a per-word confidence in a form you can act on.
Working without the page layout. If you are extracting text for a search index and you do not need the structure, the structure is overhead.
The rule
Use OCR to get the text, and use a layout-aware converter or a vision model on the pages where the structure matters. The expensive tool is for the pages that need it, not for the corpus.
The tool chain here that does this in the browser: /tools/ocr-image-to-text for the OCR, and /tools/pdf-to-text and /tools/pdf-to-images for getting there. /learn/ocr has the background and /compare/paddleocr-vs-tesseract compares the two engines.