Extract PDF Details

Extract PDF Details

Extract only information supported by the PDF and make every result easy to verify.

Workflow

1. Confirm the input PDF and infer the requested fields and output format from the user's prompt. If no fields are specified, extract a practical document summary plus the main labeled facts; do not dump all text by default.
2. Inspect the PDF before extracting:
   • Determine the page count and whether pages contain selectable text, scans, tables, forms, diagrams, or mixed content.
   • Use the available PDF-reading tools for text and layout extraction.
   • Render relevant pages when layout, images, labels, or table structure affect meaning.
   • Apply OCR to scanned or image-only pages. Inspect the rendered page when OCR is uncertain.
3. Search the whole document for each requested field, including synonyms and repeated labels. Preserve relationships between labels, values, units, variants, footnotes, and table headers.
4. Normalize the result only when safe:
   • Preserve identifiers, SKUs, model numbers, dates, units, decimal precision, capitalization, and leading zeroes.
   • Keep the original value when normalization could alter meaning.
   • Separate multiple values instead of silently choosing one.
5. Verify every extracted value against its page. Recheck nearby text, table headers, footnotes, and continuation pages.
6. Return the result in the user's requested format. If none is specified, use a compact Markdown table with Field, Value, Page, and Notes columns.

Evidence and Quality Rules

• Cite PDF page numbers for extracted facts using p. N or pp. N-M. When printed page numbers differ from PDF page indexes, state which numbering is used.
• Do not infer a value from branding, filename, visual resemblance, outside knowledge, or another document unless the user explicitly requests inference or comparison.
• Use Not found for an absent field and Unclear for an unreadable or ambiguous value. Briefly state what caused the uncertainty.
• Flag conflicts when the same field has different values in different places; report all conflicting values with their pages.
• Treat OCR output as provisional. Visually verify easily confused characters such as 0/O, 1/I/l, 5/S, punctuation, and decimal separators.
• Do not expose hidden document metadata as a document fact unless the user asks for metadata.
• Never overwrite the source PDF.

Structured Output

When JSON is requested:

• Use the user's field names as keys when possible.
• Return valid JSON without commentary inside the JSON block.
• Represent missing values as null, not invented placeholders.
• Add a separate _evidence object mapping fields to page numbers and concise notes unless the user supplies an exact schema that forbids extra keys.

When CSV or spreadsheet output is requested:

• Use one row per logical record, such as a product, invoice line, specification variant, or document.
• Repeat shared document identifiers where needed so every row remains understandable on its own.
• Keep source-page information in a dedicated column.

Completion Check

Before responding, confirm that:

• every requested field is present, Not found, or Unclear;
• every reported fact has page evidence;
• tables and multi-column text were read in the correct order;
• OCR-sensitive values were visually checked;
• units and variants remain attached to the correct values; and
• the output matches the requested schema or format.