Document AI: Common Questions
Isaac FlathSeptember 30, 2026Short answers to questions about working with documents. Watch an explanation, or follow the links for a fuller guide.
Which OCR model should I use?
Start with what your product needs from the document: plain text, tables, layout, or source locations for citations. Where the model runs and how much it costs also narrow the choices. Then test the candidates on your own documents and the task you need to accomplish.
For a fuller guide, watch Joe Barrow’s talk on choosing an OCR model. To see how I investigate failures, read Document Evals: OCR, Retrieval, and Agents.
Link to this answerWhat is a PDF, and why is it hard to OCR?
A PDF can combine stored text, scanned images, drawings, and form fields—even on the same page. It preserves how the page looks, but that does not guarantee that its reading order, table structure, or relationships are available to your software.
A scanned page needs OCR to recover its words. A page with selectable text may still need more processing to recover its meaning. The OCR Examples Gallery shows how models handle real tables, forms, charts, and diagrams.
Link to this answer