---

name: document-io-and-exports description: "Load docTR inputs and interpret or export Document, Page, KIE, table, and reading-order outputs." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

document-io-and-exports

Use this sub-skill when a task is about docTR document inputs, page arrays, result-object structure, or output conversion after OCR/KIE inference.

Route here for

• Loading PDFs, image files, image bytes, URL-rendered web pages, or already-decoded NumPy page arrays.
• Explaining DocumentFile.from_pdf, DocumentFile.from_images, DocumentFile.from_url, read_pdf, read_img_as_numpy, read_img_as_tensor, decode_img_as_tensor, or read_html.
• Interpreting Document, Page, KIEDocument, KIEPage, Block, Line, Word, Prediction, Artefact, LayoutElement, Table, and TableCell outputs.
• Exporting result objects as plain text, JSON/dicts, XML/hOCR, Markdown, AsciiDoc, or HTML with reading-order behavior.
• Diagnosing page array shape/channel/dtype, missing html or viz extras, show() dependencies, unsupported export formats, or hOCR limitations.

Do not handle here

• Creating or configuring ocr_predictor / kie_predictor, model architectures, device placement, or batch sizing; route to ../core-ocr-and-kie/SKILL.md.
• Advanced model factory, Hugging Face Hub, custom weights, ONNX export, or optimization decisions; route to ../models-and-customization/SKILL.md.
• CLI command usage or batch helper scripts; route to ../cli-and-scripts/SKILL.md.

Operating references

1. Read references/io-and-export-reference.md for input APIs, expected page array contracts, object schema, reading-order semantics, and export workflows.
2. Read references/troubleshooting.md when loading, visualization, JSON/XML, KIE, table, or format-dispatch behavior is surprising.
3. Use scripts/inspect_document_export.py for a safe local smoke inspection of DocumentFile loading and synthetic Document/KIEDocument exports. The script does not run predictors or download model weights.

Quick rules

• DocumentFile.* returns a list of pages; each page is normally a NumPy uint8 array with shape (H, W, 3) in RGB order.
• DocumentFile.from_pdf(...) rasterizes every PDF page; page index in the result corresponds to the input page order.
• DocumentFile.from_url(...) needs the html extra because it uses WeasyPrint to render a URL to PDF before rasterization.
• show() on Document, Page, KIEDocument, and KIEPage needs the viz extra because it uses visualization dependencies.
• render(), export(), export_as_xml(), Markdown, AsciiDoc, HTML, and export_as(...) share reading-order linearization by default.
• Tables are structured objects (Table/TableCell) as well as exportable text grids; table text is not duplicated in regular page blocks when table detection is used.
• OCR and KIE outputs differ: OCR pages group content spatially as blocks -> lines -> words, while KIE pages group predictions by semantic class name.