OCR a scanned PDF: accuracy, limits and privacy
Reviewed 17 November 2026 · 7 min read · Written and maintained by Toolnexo
OCR has a reputation problem: people remember the times it mangled a page, not the hundreds of times it read one perfectly. The technology is genuinely good now — within its limits. Those limits are what this guide is about, because knowing them is the difference between a usable transcript and an hour of frustration.
How our OCR pipeline actually works
Our PDF OCR tool runs on our server, and there's no hiding from that fact: Tesseract and PDF rendering are heavy native programs that don't fit in a browser tab. Each page is rendered to a 200 DPI image with pdftoppm, passed to Tesseract one page at a time, and appended to a plain-text file. The page image is deleted the moment recognition finishes.
There's a 50-page limit per file. That's not arbitrariness — a 100 MB PDF can hide thousands of pages, and rendering them all at once could exhaust the worker's storage. Split the big ones and run them in batches; it also keeps each result easier to review.
What makes OCR accurate
- Scan at 300 DPI or better. The tool renders at 200, so a sharp source gives it something to work with.
- Keep pages straight. A few degrees of skew is survivable; more and characters start smearing together.
- Contrast matters more than resolution. Grey text on a grey background is the single most common failure I've seen.
- Crop black borders and scanner shadows before running OCR.
- If a digital PDF exists, use it. OCR is only for pages that have no text layer at all.
Language, handwriting, and honest expectations
The current worker recognises English. Other languages need extra Tesseract language packs, and mixing languages on one page reliably lowers accuracy. Handwriting is a different discipline entirely — Tesseract is tuned for print, so handwritten notes come back wrong in ways that look confident. Don't trust them.
Treat every OCR result as a draft. Check names, amounts, dates and doses against the page image. A wrong digit that looks right is worse than a blank you notice immediately.
What happens to your file
The file you pick is uploaded over HTTPS, stored under a random job ID in a private volume, and deleted automatically within an hour. A "Delete now" button removes it the moment you've downloaded your result. The original filename is never used as a storage path, and downloads need a short-lived signed token — a guessed job ID gets nothing.
This is a genuine trade-off, and I'd rather state it than blur it: browser tools need no trust because nothing moves; server tools need the safeguards above. If a document is too sensitive to upload even with those safeguards, use a local OCR application — no honest web service can promise otherwise.
Checking the output
- Search for a phrase you know is on the first and last page. Missing means a page probably didn't render.
- Compare every number against the source. OCR confuses 0/O, 1/l/I, 5/S and 8/B with impressive confidence.
- Check that the page markers match the PDF's page count. A mismatch means something failed mid-run.
- Keep the original PDF next to the text file so anyone reading the transcript can verify it.
Put it into practice
PDF OCR
Turn a scanned or image-based PDF into searchable plain text — upload the PDF, get a .txt file back, with every page read by OCR on our s…
Runs on our serverPDF Text Extractor
Pull the full text layer out of a PDF page by page, with page markers, and copy or download it as a .txt file — all parsed on your device.
Runs in your browserPDF to Word
Convert a PDF into an editable Word (.docx) file on our server — upload, wait a few seconds, download the .docx.
Runs on our serverPDF Viewer
Open and read any PDF in your browser — flip through pages, jump to a page and zoom in, with the document rendered on your device and nev…
Runs in your browserOther guides
- How to merge PDFs without uploading them6 min · 17 November 2026
- Compress a PDF without destroying readability6 min · 17 November 2026
- JSON formatter: validate, format and fix common errors6 min · 17 November 2026