PDF Ingestion
Three ways to feed a PDF to the model, in increasing order of preprocessing:
-
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.
-
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. -
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
| Doc shape | Path |
|---|---|
| <20 pages, layout matters (report, form, invoice) | Native |
| <20 pages, pure prose (article, memo) | Text extraction |
| 20-100 pages, mixed | Native, but chunk if context tight |
| >100 pages | Extract → chunk → summarize |
| Scanned PDF (no text layer) | OCR first (Tesseract or vision model), then treat as extracted text |
| Tables are the point | Native — text extractors mangle tables |
| Figures/diagrams are the point | Native + explicit "describe the figure on page N" prompt |
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
pdftotextreading order. Multi-column PDFs come out as interleaved lines. Usepdftotext -layoutfor column preservation, orpdftotext -rawfor 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.
- Cache the map at a prompt-caching breakpoint so multi-turn Q&A over the doc doesn't reprocess.
Red flags
- Sending a 200-page PDF native to answer one question. Extract the relevant page range first.
- Trusting the text extractor on a form or invoice. Layout carries meaning. Use native.
- OCR'ing a PDF that already has a text layer. Check
pdftotext -q file.pdf -first — if text comes out, no OCR needed. - No page citations in output. Model can hallucinate confidently across long PDFs. Force page numbers into the response format.
- Re-uploading the same PDF every turn without caching. Cost climbs linearly; a 5-minute cache fixes it.
Loopkit-adjacent
If the PDF is a spec, extract it into PROMPT.md via spec-first — the agent should re-read prose, not re-scan the PDF, on every turn.
