PDF

PDF — task routing + recipes

Deeper material: skills/pdf/references/FULL_SKILL.md. Before ANY form filling: read src/core/tools/builtin/document/scripts/pdf/forms.md.

Route to the right tool first

Fall back to CLI/python below only for what those don't cover.

CLI recipes (qpdf + poppler, installed)

# Split: one file per page / a page range
qpdf --split-pages in.pdf out-%d.pdf
qpdf in.pdf --pages . 2-5 -- out.pdf

# Rotate all pages 90° clockwise
qpdf in.pdf out.pdf --rotate=+90

# Passwords: add / remove (needs the password)
qpdf --encrypt "$PW" "$PW" 256 -- in.pdf locked.pdf
qpdf --password="$PW" --decrypt locked.pdf out.pdf

# Embedded images (original bitmaps) vs page screenshots — different things:
pdfimages -all in.pdf img/fig     # original embedded bitmap objects
pdftoppm -png -r 300 in.pdf pg    # rasterized page renders

OCR scanned documents

Extraction returning almost no text on a scan → OCR (tesseract installed):

import pytesseract
from pdf2image import convert_from_path
text = "\n".join(pytesseract.image_to_string(p) for p in convert_from_path("scan.pdf", dpi=300))

Generating with reportlab

• Prefer Platypus (story of Paragraph/Table/Image/Spacer flowables; the engine paginates) for anything document-like; use raw Canvas only for precise placement.
• Canvas origin is BOTTOM-LEFT, y grows upward: "1 inch from the top" is page_height - inch - element_height.
• Watermark/stamp: render the stamp page with reportlab, merge onto each page with pypdf (page.merge_page(stamp)).

Always verify visually

A PDF that "generated without errors" can still be wrong. Render with document.pdf-to-images and READ the pages back (overlaps, clipped text, missing content) before delivering — mandatory after form fills and reportlab layouts.