Page numbers and bounding boxes for DOCX: filling prov in Docling's own output from a Word-compatible layout engine

[Ryujiyasu]

Docling's DOCX backend leaves prov empty, for the reason given each time it comes up (#450, #833, #997, #1757, #2865): a .docx stores no layout, so there is no page to point at. The usual advice is to print to PDF first.

I work on Oxi, an open-source (MPL-2.0) Rust/Wasm layout engine for .docx whose pagination is measured against Word's own render. I wrote a small tool that takes the DoclingDocument Docling already produces for a DOCX, plus the same file, and fills in prov and pages from Oxi's layout. Docling's structure, labels and text are left exactly as they are; nothing in Docling needs to change for this to work.

docling report.docx --to json
oxi-docling annotate report.docx report.json report.prov.json

A section header from tests/data/docx/sources/word_sample.docx after annotation (points, top-left origin, charspan over orig, one entry per page an item spans):

{
  "self_ref": "#/texts/4",
  "label": "section_header",
  "text": "Let’s swim!",
  "prov": [{"page_no": 1,
            "bbox": {"l": 66.95, "t": 380.98, "r": 167.86, "b": 405.13, "coord_origin": "TOPLEFT"},
            "charspan": [0, 11]}]
}

Checked on this repository's tests/data/docx with its ground-truth JSON as input (36 files, 845 items):

• 845 of 845 texts, tables and pictures get prov; table cells also get bbox;
• every output validates against DoclingDocument_1_10.json, with refs, pages, boxes and charspans checked;
• no file changes in anything other than prov, pages and cell bbox.

How close the pages are to Word: Oxi is scored on frozen blind sets of public documents against Word's own render (method and per-engine numbers in its README). Pagination is not perfect, and a document using a font the engine has no metrics for can paginate differently from Word. Placement rules are written down in the tool's README, including the two estimated cases (an empty paragraph the engine paints nothing for, and a comment, which takes the place of the text it annotates).

Try it:

git clone https://github.com/Ryujiyasu/oxi && cd oxi
cargo build --release -p oxidocs-docling
crates/oxidocs-docling/scripts/export_all.sh \
    <docling>/tests/data/docx/sources <docling>/tests/data/docx/groundtruth out \
    <docling-core>/docs/schemas/DoclingDocument_1_10.json

One question for the maintainers: is DOCX prov from an external producer something you would want to point users to (docs, an example), or would you rather keep DOCX provenance out of the Docling ecosystem by design?

Related: docling.rs drops prov and pages when it reads a DoclingDocument JSON (docling-project/docling.rs#899), so for now the annotated JSON should be consumed with Python docling-core.