PageIndexFrom the builder
Do We Still Need OCR?
PageIndex asks whether OCR is still needed now that vision-language models can read pages directly, arguing that flattening a 2D layout into 1D text inevitably loses spatial and contextual information and caps question-answering quality. It notes that long documents overflow the context window and that image embedding models, trained mostly on natural photos, retrieve document pages poorly. Its approach has the model reason over a generated table-of-contents tree, send the chosen page image to the vision model and repeat until it has enough, with code published on GitHub.
Converting a 2D visual document into a 1D text sequence inevitably discards spatial, contextual, and sometimes semantic relationships.
Saved . A shortened excerpt from the official blog; the original publisher retains authorship.