OCR (optical character recognition)

OCR turns pixels into characters. A classic OCR engine finds regions of text in an image and recognizes each character, producing plain text and usually a bounding box for every word. It's how scanned PDFs become searchable and how traditional document pipelines read forms.

You run into OCR as soon as documents arrive without a text layer. A scanned contract or a phone photo saved as a PDF looks normal, but text extraction returns nothing, so search and summarization silently receive empty input. Detecting image-only pages and sending them to OCR or a vision model is a basic step in any document ingestion pipeline.

The modern choice is between dedicated OCR and document-AI services and vision-capable language models. Dedicated services are strong on high-volume, fixed-template forms and return precise coordinates. Vision models understand layout and context, so they know which number is the invoice total, and they cope with endless variety, but they can misread digits or quietly normalize wording. Many teams combine them: OCR text for search and for verifying quoted values, a vision model for understanding. Either way, spot-check numbers against the source.