The full structure, not a summary

export_to_markdown() and export_to_text() both throw information away on purpose, bounding boxes, page numbers, provenance, none of that survives into a Markdown string. export_to_dict() doesn't throw anything away: it's the complete DoclingDocument structure, texts, tables, pictures, groups, and page-level metadata, as plain Python dicts and lists, ready for json.dumps().

as_dict = doc.export_to_dict()
Path("out.json").write_text(json.dumps(as_dict, indent=2))

Why this matters for a real pipeline

Running the layout, table-structure, and OCR models is the expensive part of a docling conversion, seconds per document, sometimes more. export_to_dict() is what lets you pay that cost once: convert a document, persist the JSON, and every later step (re-rendering to Markdown, pulling a table out, re-chunking with different settings) works off the cached JSON instead of re-running the pipeline.

Loading it back with DoclingDocument.model_validate()

from docling_core.types.doc.document import DoclingDocument


reloaded = DoclingDocument.model_validate(as_dict)

DoclingDocument is a Pydantic model, model_validate() is Pydantic's standard "build a model instance from a dict" method. The reloaded document is a real DoclingDocument, every method used in this course so far, export_to_markdown(), .tables[0].export_to_dataframe(), HybridChunker().chunk(), works on it identically to the object that came straight out of converter.convert().

Checkpoint

  • export_to_dict(): the complete, lossless DoclingDocument structure as plain Python data, unlike the Markdown/text exports.
  • Why cache it: the layout/table/OCR models are the expensive part of a conversion, JSON lets you pay that cost once and re-derive any other export later.
  • DoclingDocument.model_validate(d): rebuilds a fully working DoclingDocument from the dict, every export method works the same as on the original.

If anything here still feels unclear, ask before moving to Lesson 13.