Check PDF extraction before you trust the summary

Classify the PDF before extracting anything

A PDF is a page description, not a promise of clean reading order. It may contain real characters, scanned images, a hidden OCR layer, or a mixture of all three. The Library of Congress format description notes that scanned-image PDFs may not support indexing, character-based text can still extract badly when Unicode mappings are absent, and logical structure exists only when the creator includes it. A page that looks perfect can therefore produce nonsense when copied.

Open three places: the first substantive page, a dense middle page, and the last page with content. Try selecting one sentence. Search for a distinctive word visible on the page. Copy a paragraph into plain text. Record the result as native text, image-only, mixed, or uncertain. Do not run OCR over a good text layer just because OCR is available.

Test reading order, not just word recognition

Columns are the common trap. A two-column report may extract across both columns line by line, turning two valid sentences into one invalid paragraph. Headers and footers can recur inside every extracted page. Hyphenated line endings may become broken words. Compare the copied text with the rendered page and mark whether paragraphs, lists, footnotes, and section headings arrive in a usable order.

Tables need a separate test because a visually aligned grid may be stored as individually positioned text. Copy one table that includes a title, header row, a row with a blank cell, and a footnote. Rebuild it in a simple grid and compare every label and value against the page image. Never infer a blank as zero, “not applicable,” or “same as above.” If the table cannot be reconstructed confidently, cite the page and summarize the pattern qualitatively—or transcribe only the specific cells needed, with a second check.

When OCR is necessary, keep the image as the authority

Adobe’s OCR documentation says a scan initially contains image data rather than searchable text, recommends saving a backup, and says to review the result for accuracy and completeness after recognition. That last instruction matters more than the button. Choose the correct language, preserve the original, and inspect low-contrast text, small type, punctuation, decimal points, minus signs, and names. Those are the errors most likely to remain plausible.

Use a sample that reflects risk. For a short memo, review every page. For a long report, review the title page, contents, each layout type, every table used in the final work, and every quoted passage. A hypothetical quality check might compare ten randomly selected numeric cells plus every cell cited in the summary; the example is a procedure, not a claimed accuracy rate.

Gate the summary on evidence quality

Add an extraction note to the source record: method, software or manual route, pages checked, known failures, and whether the summary may rely on prose, tables, or neither. If a reviewer cannot locate a sentence on the rendered page, the extracted text is not ready for the evidence table.

• Keep the untouched PDF beside any OCR or converted copy.
• Verify every quotation against the rendered page.
• Verify table headers, units, signs, blanks, and footnotes together.
• Do not let fluent extracted text hide missing columns.

The archive’s Acrobat assistant summary describes a product claim; this guide covers the more basic control that should happen before any assistant receives the file.