docspine

A pure-Rust Word (.docx) parser with Python bindings (PyO3 / maturin, abi3-py311). A .docx file is OOXML — a zip archive of XML parts — and docspine walks word/document.xml directly to produce a structured, information-preserving model: paragraphs (styled runs), tables (rows, cells, merges, fills, nesting), and embedded pictures. Tables are a first-class focus. Embedded images can additionally be OCR'd locally, offline, and deterministically via the sibling ocrspine crate (PP-OCRv5 through tract-onnx — no cloud, no network), and an image that is a table can be reconstructed into a grid from its OCR word boxes. A parsed document can also be exported to PDF (to_pdf() / save_pdf()) with flowed layout and pagination through the shared pure-Rust pdf-typeset engine from pdfspine — no LibreOffice, no cloud converter.

docspine is the document-engine sibling of pdfspine (PDF) and pptspine (PowerPoint), all sharing the same ocrspine OCR core.

Capabilities

Parsing is tolerant: unknown elements are skipped, missing attributes become None, and malformed input yields a typed DocError rather than a panic.

docx first; legacy .doc deferred

Modern .docx (OOXML) is the primary target. The old binary .doc is a Microsoft compound document (OLE/CFB): rebuilding its body from the binary FIB + piece table is large, fiddly, and shares almost nothing with the docx path. So docspine ships detection + a clean typed downgrade today (a .doc byte stream yields DocUnsupportedError, and probe_doc reports the CFB streams when built with the legacy-doc feature); full .doc body reconstruction is a follow-up, not a blocker.

Install

pip install docspine

docspine is on PyPI. OCR works out of the box: the PP-OCRv5 weights ship in the shared ocrspine-models data package — a runtime dependency pip pulls in automatically — so a plain pip install docspine finds the OCR models with no extra setup (it no longer needs a sibling ../ocrspine/models checkout or OCRSPINE_MODELS). To build from source instead, see below.

Build (from the package root)

uv venv .venv
VIRTUAL_ENV="$(pwd)/.venv" uv pip install maturin pytest
# Structural parsing needs no models. The OCR path resolves models from a
# sibling ../ocrspine/models by default (or set OCRSPINE_MODELS).
OCRSPINE_MODELS="$(cd ../ocrspine && pwd)/models" \
  VIRTUAL_ENV="$(pwd)/.venv" .venv/bin/maturin develop --release

Use from Python

import docspine

doc = docspine.open("report.docx")
print(doc.block_count)

for block in doc.body():            # list[dict], introspectable
    if block["kind"] == "paragraph":
        for run in block["runs"]:
            print(run["text"], run["bold"], run["color"])
    elif block["kind"] == "table":
        for row in block["rows"]:
            for cell in row["cells"]:
                print(cell["text"], "span", cell["grid_span"], cell["v_merge"])

# Run OCR on raw image bytes (PNG/JPEG), offline:
items = docspine.ocr_image(open("scan.png", "rb").read())
print(" ".join(i["text"] for i in items))

# Reconstruct a table that lives inside an image into a grid:
for table in docspine.reconstruct_image_table(open("table.png", "rb").read()):
    for cell in table["cells"]:
        print(cell["row"], cell["col"], cell["text"])

Export to PDF

doc = docspine.open("report.docx")
doc.save_pdf("report.pdf")         # flowed layout + pagination
pdf_bytes = doc.to_pdf()           # or in-memory bytes

# Optional: map a requested font family to a local font file (or to another
# installed family), layered on top of the built-in substitution table:
doc.save_pdf("report.pdf", font_map={"Calibri": "/path/to/Carlito.ttf"})

Per-section page geometry from sectPr is honored (paper size, orientation, margins per section). Rendering is deterministic and fully offline. Missing fonts degrade gracefully: an available face is substituted and a Python UserWarning is emitted once per warning kind — the export never fails on a missing font.

Rust workspace

crates/
  doc-core    domain model + geometry (twip/EMU) + typed DocError. No IO/zip/XML.
  doc-parse   OOXML reader: zip extract + quick-xml walk -> Document.
              Legacy binary .doc probing behind the `legacy-doc` feature.
  doc-ocr     image-OCR bridge over ocrspine (PaddleOcr) + image-table reconstruction.
  doc-render  docx -> PDF renderer over the shared pdf-typeset engine (from pdfspine).
  py-bindings PyO3 _core extension (the FFI chokepoint); `ocr` feature gates OCR.

Deferred / follow-up

• Full legacy binary .doc (OLE/CFB / [MS-DOC]) body reconstruction (FIB, piece table, CHPX/PAPX). Today: detection + typed downgrade + probe_doc.
• Richer styling, headers/footers, footnotes/endnotes, comments, fields, hyperlinks targets, SmartArt/charts.