Vision and document understanding

Vision-capable models read images, charts, scans and PDFs directly, often replacing classic OCR pipelines.

Evolving: the basics hold, details shift. Reviewed 17 Oct 2026.

Key ideas

• Send PDFs page-as-image plus extracted text for best results on tables and layouts.
• For high volume, use a parser (Docling, Unstructured, Textract, Azure Document Intelligence) and reserve vision models for hard pages.
• Ask for structured output with field-level confidence or source page numbers.
• Image tokens are expensive; downscale to the resolution the task needs.