Skip to content

Parse any format

parse() accepts PDF, DOCX, PPTX, and single images (PNG/JPEG/WebP), and produces the same layout-aware ParseResult for all of them. Requires the OCR server.

PDFs

from ingestlib.operations import parse

result = parse("report.pdf")
print(result.page_count, "pages in", result.parse_duration_seconds, "s")

Pages render at 200 dpi by default — a balance of OCR accuracy against VLM token cost. Raise it for dense small print:

result = parse("dense-datasheet.pdf", dpi=300)

Scanned PDFs need nothing special: parse always OCRs the page render and never trusts the file's embedded text layer.

Password-protected or corrupt PDFs

Both fail immediately with a plain-language error — remove the password (print/export to a new PDF), or re-export the corrupt file.

Word & PowerPoint

result = parse("slides.pptx")
print(result.was_converted)   # True — went through PDF on the way

DOCX/PPTX convert to PDF via LibreOffice first. If LibreOffice isn't installed, the error says exactly what to install — PDFs and images keep working without it.

Images

A single PNG, JPEG, or WebP parses as a one-page document:

result = parse("whiteboard-photo.png")
page = result.pages[0]
print(page.markdown)

Same OCR, same chart reading, same regions and bounding boxes — an image is simply a page with no text layer.

Async

Inside an event loop (a notebook, a web handler), use the a-prefixed form — the sync form will tell you so if you forget:

from ingestlib.operations import aparse

result = await aparse("report.pdf")

Cost & performance notes

  • Parse is the expensive stage: OCR per page plus LLM enrichment for charts/figures plus a review pass — roughly $0.002/page in LLM spend on the default stack.
  • The same file never parses twice in the pipeline: ingest() dedups by content checksum. Calling parse() directly, reuse the returned ParseResult object — feed it straight to classify/split/extract, no re-parse.

Next: Use parse results — markdown, tables, figures, and page images.