xuzhougeng Read PDFs page-by-page, not wholesale
read chokes on PDF binary, and pasting a 50-page document costs 40K+
tokens. The sidecar parses once into the persistent Python kernel (memory +
disk cached), after which you pull exactly the pages the question needs.
Setup, once per session: run the exec(...) line from the "Python
Runtime Sidecar" section at the end of this skill's use_skill output.
Definitions survive across cells until the kernel restarts. pypdfium2 is
required (pillow too for image mode); if the first call raises
ImportError, follow its hint and re-run.
Pick the entry point
| call | use for | gives |
|---|---|---|
pdf_outline(path) |
any structured document — start here | [{page, heading, level}] from embedded bookmarks, [] + hint when absent |
pdf_pages(path, pages=[...], mode="text") |
the specific pages you need | [{page, text, n_chars}] |
pdf_pages(path, mode="image", dpi=200, pages=[N]) |
figures, scans | one PNG per page in .cache/pdf-explore/, for view_image |
default mode="auto" |
unknown file | text, auto-switching to images when pages have no text layer |
Map the document first
toc = pdf_outline("report.pdf")
for entry in toc:
indent = " " * (entry["level"] - 1)
print(f'p{entry["page"]:>3} {indent}{entry["heading"]}')
Costs nothing when bookmarks exist (LaTeX-compiled papers almost always
have them). On [], there is no LLM fallback here — print the opening
lines of each page from pdf_pages(path, mode="text") and build the map
yourself. Watch for the [pdf_outline] offset warning: some PDFs bookmark
logical page numbers, which are shifted from file page numbers by the
front matter.
A handful of pages: print them
hits = pdf_pages("report.pdf", pages=[12, 13], mode="text")
for h in hits:
print(f'\n[page {h["page"]}]\n{h["text"]}')
Fine up to roughly five pages (~2–4KB each). Kernel output past the ~16KB context budget is head/tail-truncated at ingestion, so anything larger goes through a file instead.
Whole sections: go through a file
For "summarize the methods", cross-section comparisons, or any multi-range
pull, write all wanted pages in one call and read the result — read
output enters context untruncated:
section_pages = [5, *range(21, 26), 62, 63, 64] # from the outline
chunks = pdf_pages("report.pdf", pages=section_pages, mode="text")
open("pull.txt", "w").write(
"".join(f'\n[page {c["page"]}]\n{c["text"]}' for c in chunks))
print("bytes:", __import__("os").path.getsize("pull.txt"))
Then read pull.txt, with offset/limit when it's long. As text a
page runs ~800 tokens; as an attached image ~8K — and the parse is paid
once.
Figures: render high, crop tight
A whole-page render can't resolve axis labels on a dense figure. Render at high dpi, crop to the figure with PIL, and view the crop:
page = pdf_pages("report.pdf", mode="image", pages=[7], dpi=200)[0]
from PIL import Image
Image.open(page["image_path"]).crop((x0, y0, x1, y1)).save("panel7.png")
view_image the crop (or the full image_path once, to locate the
figure). Every viewed image stays in context until /compact ages it out —
view the few crops that matter, never the whole render set.
Boundaries
The reference host's LLM helpers (pdf_scan page ranking, pdf_extract
sweeps, pdf_map per-page summaries) require an in-kernel model bridge
Wisp doesn't provide, so they don't exist here. For an exhaustive pass,
dump pages to files in chunks (recipe above) and work through them, or hand
the on-disk text to the explore subagent.