This demo runs on Transformers, not the SGLang serving stack, so it does not do unlimited-horizon decoding and is intended for small-scale testing only. To parse genuinely long documents, deploy the model with SGLang — see GitHub or the Hugging Face model card. To keep output stable a request is capped at 20 pages; longer documents are truncated to their first 20 pages before parsing.
One request, whole document. Every page you add goes into a single request, so cross-page context is available while parsing.
Rendered view. Each block is rendered by its detected type — titles, tables, equations — figures are cropped out of the page image, and blocks split across a page break are merged back together.