AI Document Reading: Text, Tables, Footnotes and Images

Before using an AI agent’s account of a document, check the representation it read. A text extraction can contain the headline and every number while losing which table heading applies to each number. A page image can preserve the layout while leaving small footnotes unreadable at the supplied resolution.

This reference is for editors and builders who need to know whether a source supports an extracted passage. Its sixteen selected checks concern information inside a document, especially PDFs. They do not measure any model’s reading accuracy. The useful question is whether the specific source feature survived the path from document to agent.

Use the table on the pages that matter to the decision. Text layer means machine-readable characters stored in the file. OCR, or optical character recognition, creates text from an image. Neither term promises that paragraph order, table associations or small print survived correctly. The final column identifies what to compare before using the passage.

W3C PDF3 explains the relationship between logical reading order and a tagged PDF’s structure. PDF6 addresses table markup that preserves row and column relationships. Those are accessibility techniques, not tests of an AI model, but they identify information a flat text dump can fail to communicate.

W3C PDF7 distinguishes scanned images of text from actual text and includes checking OCR output. PDF1 explains text alternatives that convey an image’s meaning. An alternative description can help interpretation without being a complete numerical transcription of a chart.

Our table applies these distinctions to agent reading. It does not claim WCAG conformance, universal PDF behavior or that a particular tool uses document tags. Inspect the tool’s actual output. If a renderer supplies only page images, the existence of tags in the original does not establish that the agent received them.

Consider a hypothetical report with columns for current customers and new customers. A footnote excludes trials from only one column. An extraction containing both totals and the footnote text may still join the exclusion to the wrong population. The numbers are present; the relationship is the missing evidence.

Choose a claim-bearing cell and record its row label, every applicable column heading, unit and note marker. Compare that record with the rendered page. If the table continues over a page boundary, inspect both pages. A repeated header can be mistaken for data, while a missing header can leave later values detached from their meaning.

Do not infer that all other cells passed because one cell did. State the scope of the check. For a decision depending on the whole table, inspect the complete relevant table or obtain a source format that preserves its structure.

Attach a short record to the extraction: file and revision, page range supplied, representation used, sections inspected, passages unresolved and any external attachments not opened. The record should make omissions visible. “Read the report” is too broad if the tool returned selected search snippets or only the first pages.

A footnote marker without its note is a useful stop signal. Look for the matching note on the page, at the end of the section or in an appendix. If it cannot be located, retain the narrower claim that the available text supports. Do not invent an exclusion because a familiar reporting convention would normally use one.

Our citation-checking reference begins once the relevant source passage is available. This document check comes earlier: it asks whether the agent had the necessary passage and relationships to assess the claim at all.

If words are missing, inspect the image or rerun extraction with a suitable method. If relationships are missing, recover the table or page structure. If a chart’s values cannot be read accurately, look for its accompanying data rather than estimating a precise number from a small image.

Keep extracted text and your correction distinguishable. In a hypothetical note, “footnote manually checked on page 12” tells the next reviewer more than silently inserting a qualifier. The source-independence reference explains why these different representations remain views of one evidence origin.

For audio-derived material, use the separate transcript verification reference. Speaker attribution and missing sound require checks this PDF-oriented table does not cover. A combined research task can need both records without pretending that one checklist proves the entire source set.