PDF work with Python

Runtime

The active Context has an isolated project in ./agent. Add packages for the operation at hand:

uv --directory ./agent add pypdf
uv --directory ./agent add pymupdf
uv --directory ./agent add reportlab

Use uv --directory ./agent run python ... for scripts. Resolve script, input, and output paths to absolute paths before calling uv, because --directory changes the command's working directory. Write working files to the current Context's ./playground/<scope>/ directory and publish the final PDF through the available delivery tool.

The Runtime normally provides poppler, qpdf, and tesseract in ./agent/conda; use micromamba run --prefix ./agent/conda <command> when one is needed. Add a missing native tool with micromamba install --yes --override-channels --channel conda-forge --prefix ./agent/conda <package>. Do not use apt, pip, conda, or mamba for setup.

Choose the operation

Check whether text extraction works before starting OCR. Preserve source page order, orientation, crop boxes, annotations, and form values unless the task asks to change them. For protected documents, respect the user's access and the PDF's permissions.

Small page operation

from pypdf import PdfReader, PdfWriter

reader = PdfReader("/absolute/path/input.pdf")
writer = PdfWriter()
for page in reader.pages:
    writer.add_page(page)
with open("/absolute/path/output.pdf", "wb") as file:
    writer.write(file)

Use document metadata supplied by the user or derived from the document. Do not insert a vendor name, web address, or promotional metadata.

Check the result

Open the result again with pypdf or pymupdf. Compare page count, page sizes, text or form fields, and metadata with the request. Render and visually inspect every affected page for clipping, missing glyphs, and misplaced overlays. For generated CJK PDFs, register and embed a font that covers the document's characters; verify extracted text as well as the render. If a needed font or renderer is unavailable, describe the remaining check plainly.