Word conversion script check

Checked on 2026-10-14 with two synthetic DOCX files. This checks the bundled Python converter, not an end-to-end Agent session or security certification.

Version and environment

Windows, Python 3.12.14, MarkItDown 0.1.8. The upstream requirement is markitdown[docx]>=0.1.0, a lower bound rather than an exact version lock.

Pinned source: GitHub Awesome Copilot d7e4ad98. Script SHA-256: b2478a03e2939f9d826143856f2595d57392d6b1d624ac8612e417f95b34ff72.

Input and output

sample.docx contained a Heading 1, two bullet items, a two-row Metric/Amount table with fictional Revenue/125 and Cost/40, and blue/red/blue PNG images. text-only.docx contained a Heading 1 and one paragraph with no pictures.

The single-file run produced sample/sample.md beside the input. A batch run with an explicit fresh output parent converted 2/2 files. Both returned exit code 0. The output retained the heading, bullet text, table labels, and numeric values.

Markdown image links referenced img/img001.png, img/img002.png, and img/img003.png in the original blue/red/blue order. Each file matched the corresponding source PNG bytes, and no data:image placeholder remained. The text-only output had no img directory. Input DOCX SHA-256 values were unchanged.

A missing input path and an existing unsupported .txt file each returned exit code 3.

What was not checked

Copilot loading, installation, recursive folders, complex tables, pagination, comments, tracked changes, protected documents, customer files, and security were not tested. Word visual rendering was unavailable because the environment lacked LibreOffice; no layout-fidelity conclusion is made.

Source inspection found that convert_one deletes an existing destination directory before writing. All tested destinations were new and confined to the test folder; overwrite deletion was not exercised. Back up manually edited output before rerunning.

This result does not establish accuracy for arbitrary documents or a traffic improvement.