DOCX Processing Openai

DOCX reading, creation, and review guidance

Reading DOCXs

• Use soffice -env:UserInstallation=file:///tmp/lo_profile_$$ --headless --convert-to pdf --outdir $OUTDIR $INPUT_DOCX to convert DOCXs to PDFs.
  • The -env:UserInstallation=file:///tmp/lo_profile_$$ flag is important. Otherwise, it will time out.
• Then Convert the PDF to page images so you can visually inspect the result:
  • pdftoppm -png $OUTDIR/$BASENAME.pdf $OUTDIR/$BASENAME
• Then open the PNGs and read the images.
• Only do python printing as a last resort because you will miss important details with text extraction (e.g. figures, tables, diagrams).

Primary tooling for creating DOCXs

• Create and edit DOCX files with python-docx. Use it to control structure, styles, tables, and lists. Install it with pip install python-docx if it's not already installed.
• After every meaningful batch of edits—new sections, layout tweaks, styling changes—render the DOCX to PDF:
  • soffice -env:UserInstallation=file:///tmp/lo_profile_$$ --headless --convert-to pdf --outdir $OUTDIR $INPUT_DOCX
• Convert the PDF to page images so you can visually inspect the result:
  • pdftoppm -png $OUTDIR/$BASENAME.pdf $OUTDIR/$BASENAME
• Inspect every PNG before moving on. If you see any defect, fix the DOCX and repeat the render → inspect loop until all pages look perfect.

Quality expectations

• Aim for a client-ready document: consistent typography, spacing, margins, and layout hierarchy. Heading levels should be obvious, lists aligned, and paragraphs easy to scan.
• Never ship obvious formatting defects such as clipped or overlapping text, default-template styling, broken tables, unreadable characters, or inconsistent bullet styling.
• Charts, tables, and visuals must be legible in the rendered PNGs—no pixelation, misalignment, missing labels, or mismatched colors.
• Never use the U+2011 non-breaking hyphen or other unicode dashes as they will not be rendered correctly. Use ASCII hyphens instead.
• Citations, references, and footnotes must be human-readable and professional. No tool-internal tokens (e.g., [145036110387964†L158-L160]), malformed URLs, or placeholder text should be present in the document.
• You must convert all citations into a human-readable format in the document with standard scholarly citation format. No 【【turn1541736113682297662view0†L11-L19】 notations are allowed in the document as the reader cannot interpret them (such citations will be severely penalized).
• Content should be concise, relevant, and free of boilerplate AI phrasing. Ensure each section adds value and flows logically.

Final checks

• Re-run the DOCX → PDF → PNG loop after your final changes and inspect every page at 100% zoom. Look for subtle issues like inconsistent spacing, widows/orphans, or misaligned bullet levels.
• Correct every formatting defect you see in the PNGs, including but not limited to: overlapping text or shapes, clipped text or shapes that are cut off, black squares, broken tables, unreadable characters, etc.
• Only deliver the DOCX once the latest PNG review confirms the document is visually flawless and professionally styled.
• Keep intermediate files organized (or cleaned up) so reviewers can easily locate final outputs.