A PDF is not just a long text file. The answer might live in a footnote, a chart legend, a table column, a page header, or the relation between a diagram and its caption. Sending a 200-page report to an AI with “summarize this” can produce a fluent overview, but it gives neither you nor the model a clear way to distinguish a page that merely mentions a topic from the page that proves the answer.
The practical move is to locate before you read. Turn a document question into two bounded passes. First, find the few candidate pages and name what each one might support. Then inspect only those pages—with the visual detail the claim needs—and return an answer tied to page-level evidence. This is not a substitute for reading a short, simple PDF normally. It is a way to make large or visually dense documents answerable without treating every page as equally relevant.
Modern file inputs can preserve both parts of a PDF. OpenAI says its vision-capable PDF flow supplies extracted text and page images to the model, and offers low or high image detail for PDF pages.[1] Anthropic describes a similar approach: it extracts page text, renders each page as an image, and lets the model analyze both text and visual content such as charts, tables, and diagrams.[2] That capability is useful only when the workflow asks a page-specific question instead of one broad request that makes every page compete for attention.
Separate page discovery from evidence inspection
Start with the decision you need to make. “Find the reported operating margin for 2025” needs a number, period, unit, entity, and an authoritative table or note. “What does the architecture diagram require?” needs the diagram, its labels, and perhaps surrounding explanatory text. Those are different evidence shapes, so do not begin with the generic instruction to summarize.
Pass one is discovery. Ask for a compact page index, not an answer. Use headings, page numbers, keywords, table titles, captions, and visible section labels to return at most a small set of candidates. Each candidate should say what it may contain and whether the evidence is textual, tabular, graphical, or mixed.
Pass two is inspection. Re-send or focus on the named pages. Ask the model to extract the exact claim, the page number, the visual or textual locator, and any qualifier that changes the interpretation. For a chart, require the series, axes, units, period, and legend—not just the tallest-looking bar. For a table, require the row label, column label, unit, and any note marker. The result is a claim you can check in the original page rather than a remembered paragraph from the whole file.
Copy this PDF question card
Use this for annual reports, policy manuals, technical specifications, slide exports, or scanned packets. Replace the brackets; keep the pass-one result small.
QUESTION
[The exact question to answer]
PASS 1 — LOCATE (no final answer yet)
- Document: [title/version/file ID]
- Find at most [3–5] pages that could directly answer QUESTION.
- For each candidate, return:
{page, section_or_visible_heading, evidence_type: text|table|chart|diagram|mixed,
supports: [specific part of QUESTION], locator: [table title/caption/keyword],
uncertainty: [what this page may still not establish]}
- If no page directly supports a required part, say `not located`.
PASS 2 — INSPECT THE NAMED PAGES
- Inspect only candidate pages [page numbers].
- Return:
{answer_or_bounded_unknown, page, exact_locator, evidence: [short quote or
faithful table/chart transcription], qualifiers: [unit/date/scope/note],
confidence_limit: [what the page does not prove]}
- Do not infer a number, category, or relationship that is absent from the page.
If your API exposes a detail control, spend high visual detail selectively. OpenAI documents high detail as the option for dense charts, small print, and diagrams, while low detail uses fewer input tokens.[1] Anthropic warns that dense pages, complex tables, and heavy graphics can fill a context window before a document reaches its page limit, and recommends splitting large documents into sections when necessary.[2] The operational rule is simple: use the first pass to earn the expensive second pass.
Preserve the page reference in the final handoff
A page number alone is sometimes fragile: PDF page labels can differ from the viewer’s count, and tables may continue across pages. Keep the document ID or version, page number as shown by the viewer, section heading, and one local locator such as “Table 7, row Operating margin, column FY2025.” If the document is mutable or consequential, preserve the version you inspected before a decision is made.
This also makes review fast. A teammate does not need to reread the report; they can open the named page and see whether the row, chart, or quote supports the claim. If the first pass finds two plausible pages with different periods or definitions, do not average them into one answer. Inspect both and state the distinction.
Failure mode: mistaking visual access for verified extraction
A PDF-aware model can still misread a tiny footnote, confuse similar columns, miss a chart’s axis scale, or answer from an adjacent period. Rendering pages is not proof that the number was extracted correctly. For decisions that affect money, compliance, health, or a public claim, have a person compare the final extraction with the named page. Keep the answer bounded when the scan is illegible, the table is ambiguous, or a required qualifier is missing.
The habit to keep
do not ask AI to absorb an entire PDF before it knows where the evidence lives. First locate the few pages that could answer the question; then inspect the visual and textual details that make the answer checkable.