A scanned PDF is a photograph of a page in a PDF wrapper. There is no text in it at all, which is why you cannot select anything, search finds nothing, and an ordinary extractor hands back an empty box. Getting the words out means recognising them, and this does that in your browser: every page is rendered at roughly 300 DPI — the resolution where recognition accuracy stops improving — and read by Tesseract compiled to WebAssembly. Nothing is uploaded, which is the whole point, because the documents that arrive as scans are contracts, medical records, court filings and bank statements, and those are exactly the files people should not be handing to a free web service. Fifteen languages are available, and choosing only the ones present matters more than it sounds: every extra language gives the recogniser more characters to confuse. Text comes back per file in an editable box with wrapped lines rejoined into paragraphs, and with a confidence figure, because OCR fails by inventing plausible words rather than by leaving gaps.
| Source | Expected result |
|---|---|
| 300 DPI scan of printed text | Excellent — typically above 95% |
| 150 DPI scan | Good, with more errors in small type |
| Photocopy of a photocopy | Mixed — speckling confuses the recogniser |
| Photographed page | Good if square, sharp and evenly lit |
| Skewed or curled pages | Noticeably worse — straighten first |
| Digitally-created PDF | Use the text extractor instead — it is exact |
| Handwriting | Poor — Tesseract is trained on printed type |
No. Pages are rendered and recognised in your browser. The language model comes from a public CDN once and is cached; your document never leaves the device. That is the reason to use this rather than a server-based converter — scanned documents are disproportionately contracts, medical records and bank statements.
An extractor reads a text layer that already exists. A scanned PDF has none — it is an image — so an extractor returns nothing. This recognises the characters from the picture instead. If your PDF was made digitally, the extractor is faster and perfectly accurate, so try that first.
Every page is rendered at 300 DPI and then recognised, which is genuinely heavy work, and it is all happening on your machine rather than a server farm. Expect a few seconds per page. The first run also downloads the language model.
On a clean 300 DPI scan of printed text, typically well above 95 percent. Accuracy falls sharply with skew, low contrast, speckled photocopies and small type. The confidence figure reported alongside the text tells you which case you are in.
It is Tesseract's own average certainty. Below 80 percent, proofread numbers and names carefully; below 60, treat the output as a rough draft. This matters because OCR does not leave blanks where it struggles — it produces confident-looking wrong words.
Yes, but it processes one page at a time and each takes a few seconds, so a few hundred pages will take a while. Each page's bitmap is released as soon as it is read, so memory stays flat however long the document is.
No. You get the text in reading order, not a reconstruction of columns or tables. Rebuilding layout from a scan is a substantially harder problem than recognising the characters.
A scanned PDF is a photograph of a page wrapped in a PDF container: there is no text layer to extract, which is why copy and search both fail on it and why an ordinary text extractor returns nothing. Recovering the words requires actually recognising them. Each page is rendered to a canvas with pdf.js at roughly 300 DPI — a PDF page is defined at 72 DPI, so the render scale is a little over four — because that is the resolution at which Tesseract's accuracy stops improving; rendering larger costs time and memory and buys nothing. The canvas is handed to Tesseract compiled to WebAssembly, and released immediately afterwards, since a 300 DPI A4 bitmap is around 25 megabytes and holding every page of a long scan at once would exhaust the tab. One worker is created per language selection and reused across every page and every file, because initialising a worker re-downloads and re-parses 10 to 15 megabytes of training data, which dwarfs the cost of recognising another page. Raw output is then tidied: Tesseract breaks lines where the scan did, so wrapped lines are rejoined into paragraphs, words split by a trailing hyphen are repaired without leaving the hyphen, and the form feeds Tesseract emits between pages become paragraph breaks. Confidence is averaged across pages and reported, with a warning below eighty percent — a low-confidence result that still reads like prose is more dangerous than an empty one, because nothing about it signals that a figure or a name may have been invented. The entire pipeline runs in the page, so confidential scans are never uploaded.