PDF Ingestion

PDF Ingestion

Three ways to feed a PDF to the model, in increasing order of preprocessing:

1. Native PDF input — pass the file directly. Model sees pages as images + extracted text. Best for docs under ~100 pages with meaningful layout (tables, figures, forms). Preserves structure.

2. Text extraction then send — pdftotext / pypdf / equivalent, then send the text. Loses layout but cheap. Fine for prose-heavy docs where tables don't matter.

3. Extract → chunk → summarize → send — for docs >100 pages or when you'll query the same doc many times. Preprocess once, cache the summary.

Deciding which path

Native PDF — the good defaults

• Cache the PDF at a prompt-caching breakpoint (see prompt-caching). Native PDFs are large — every uncached turn costs full input price on the whole doc.
• Ask about specific pages ("summarize section 3.2 on page 14") rather than the whole doc. The model handles targeted queries better than "summarize this 80-page report".
• Follow up with page-cited claims — "on which page does the doc say X?" — as a sanity check the model isn't hallucinating.

Extract-then-send — the traps

• pdftotext reading order. Multi-column PDFs come out as interleaved lines. Use pdftotext -layout for column preservation, or pdftotext -raw for straight reading order — pick per doc, don't guess.
• Tables become word soup. If tables are load-bearing, native or per-table image extraction. Not text.
• Headers/footers repeat on every page. Strip them before sending, or the model will treat them as content.
• Footnotes drift to random positions in the extracted stream. Filter or accept the noise.

Extract → chunk → summarize (long docs)

• Chunk by section, not by token count. A section-aware split respects the doc's logic; a naive 4K-token split cuts sentences and tables.
• Summarize per section into a "map" — 1-2 paragraphs each. Keep the map short enough to fit in context whole (~2-4K tokens for a 200-page doc).
• Store the full section text alongside the map (paths in a manifest). Fetch on demand when a question needs detail beyond the summary.