OCR

OCR stands for Optical Character Recognition. A scanned document, a photo of an invoice or a PDF from a photocopier is just an image: the computer sees pixels, not letters. OCR analyses that image, identifies the characters and returns the text in a usable form.

The first OCR software compared the shape of letters with stored templates. Today's tools rely on neural networks that read whole lines and handle several languages, tables, page layout and, to some extent, handwriting. Recent vision models go further: they understand the structure of a document and can directly extract fields such as a date, an amount or a contract number. The result is never perfect, especially on a blurry or skewed scan, which is why clean scanning is worthwhile.

OCR is a key step before RAG: an AI assistant cannot cite a document it cannot read. If your archives contain many scanned PDFs, plan for this conversion at indexing time. Classic OCR can run on the processor, but a GPU greatly speeds up processing of large volumes, especially with vision models, which then need VRAM in addition to that used by the LLM.

When we install a private AI assistant, we work with you on processing your scanned documents so that they too become searchable.