Does this work on scanned PDFs?

Yes, and that is the ordinary case. Extraction never reads the PDF's text layer — every page is read from its pixels — so a scan without a text layer works the same way as anything else.

Scans in 15 scripts, each checked against a real scanned page in one language: Arabic, Bengali, Chinese, English, Greek, Hebrew, Hindi, Japanese, Kannada, Korean, Malayalam, Russian, Tamil, Telugu, and Thai. Other languages written in those scripts go through the same model, but only these were tested. Traditional Chinese is read with the Simplified model.

How big a PDF can I save?

Up to 300 pages and 500 MB. Past either limit the file is rejected outright.

How do I save a PDF to read later?

Open it in Chrome or Firefox and click the Readplace extension, or share it from Safari on an iPhone. Either way it lands in your readlist with your other articles. Pasting a link on this page opens it in the reader without saving it.

Can I read an extracted PDF on a Kindle or Kobo?

Yes. The reader has a Download EPUB button, and the EPUB is built from the extracted text, so it reflows on e-ink too.

Do I need an account to try it?

No. Paste a PDF link into the reader and read the result.

What does it cost to keep using it?

Making an account starts a 14-day trial of the full product with no card asked for. After that you pick how often you pay: $10 billed monthly, $60 billed once a year, or $108 billed once every 3 years. Every one of them is the same whole product.

What happens to PDFs I already saved if I stop paying?

You keep reading them. The account goes read-only: the readlist, the reader view, the extracted text and export all keep working. Saving new links and importing are what stop.

For readers who save papers, reports and scans

Read the PDF, not a guess at it.

Paste a PDF link to read it now, with no account. To keep PDFs, save them the way you save articles, from the browser extension on the open tab or the iPhone share sheet, and each one lands in your readlist as text. Scans are read by OCR from their pixels, and a cleanup pass that changes any number is thrown away.

Paste a PDF link and read the result. No account, no download.

How a PDF becomes text

Three stages, and each one can be thrown away.

1

Every page is rasterised

Each page is rendered to a 300 DPI image and read by Tesseract. This happens whether or not the PDF claims to have a text layer, so a scan with no text layer is the ordinary case rather than a special one.
2

Three passes clean it up

One language-model pass fixes OCR noise, a second reviews a word-level diff of those edits, and a third turns the result into structured HTML.
3

Each pass has to survive a check

The two passes that touch the words are checked against the raw Tesseract text; the pass that turns it into HTML is checked for text it dropped. Fail either and that pass is discarded and the stage below it is kept, so a rejected rewrite never reaches you.

Where an extracted PDF ends up

The Readplace reader showing an article with its AI summary expanded — The reader view any saved link opens into. An extracted PDF lands here too — as text that reflows on a phone, not as a page you pinch and drag.

Fayner Brack, Founder & CEO of Readplace, wrote js-cookie, which browsers download about 22 billion times a year, and ran his own reading pipeline for ten years before it became Readplace. Why Readplace exists.

What the checks actually check

This is the part worth being precise about, because a language model in a pipeline usually means the opposite.

Three deterministic checks run against the raw Tesseract output after each of the two passes that touch the words: the same runs of digits must all come back, the total length must stay within 30 percent, and the line and blank-line structure must be unchanged. The third pass only turns text into HTML, so it is checked for how much text it dropped rather than for what it said.

The digit check is the one that earns its place. Dates, page numbers, citations, figures and table values are what a language model is most likely to quietly alter, and what a reader is least likely to catch. If they change, the pass is dropped.

This is not a guarantee that no word ever changes. A same-length substitution of one non-numeric word for another passes all three checks. The narrower claim is the true one: altered numbers are caught, and a rejected rewrite is discarded rather than shipped.

What this does not do

Scans are read in 15 scripts, each tested with one language: Arabic, Bengali, Chinese, English, Greek, Hebrew, Hindi, Japanese, Kannada, Korean, Malayalam, Russian, Tamil, Telugu, and Thai. A page in another script, such as Georgian, can be read with the wrong one and come back as noise.
PDFs up to 300 pages and 500 MB. Past either limit the file is rejected rather than partly processed.
If more than 20 percent of pages fail OCR the whole extraction is rejected. Below that, failed pages appear as OCR-unavailable markers.
Non-numeric words can still change within the length and structure bounds. The checks catch altered numbers, not every altered word.
Every page is re-rasterised even when the PDF has a clean text layer, which costs time that reading the text layer would have saved.
Extraction runs after the page opens, so a long PDF takes a few minutes to fill in.
A PDF with no web address behind it, such as one shared straight from the iPhone Files app, can't be saved, because Readplace files everything by its link.

Questions

Does this work on scanned PDFs?: Yes, and that is the ordinary case. Extraction never reads the PDF's text layer — every page is read from its pixels — so a scan without a text layer works the same way as anything else.
Can the language model make things up?: It can change words, and the checks are built around that. On the two passes that touch the words, the same runs of digits must all come back, the total length must stay within 30 percent, and the line structure must match. A pass that fails any of those is discarded and the rawer text underneath is kept.
What languages work?: Scans in 15 scripts, each checked against a real scanned page in one language: Arabic, Bengali, Chinese, English, Greek, Hebrew, Hindi, Japanese, Kannada, Korean, Malayalam, Russian, Tamil, Telugu, and Thai. Other languages written in those scripts go through the same model, but only these were tested. Traditional Chinese is read with the Simplified model.
How big a PDF can I save?: Up to 300 pages and 500 MB. Past either limit the file is rejected outright.
How do I save a PDF to read later?: Open it in Chrome or Firefox and click the Readplace extension, or share it from Safari on an iPhone. Either way it lands in your readlist with your other articles. Pasting a link on this page opens it in the reader without saving it.
Can I read an extracted PDF on a Kindle or Kobo?: Yes. The reader has a Download EPUB button, and the EPUB is built from the extracted text, so it reflows on e-ink too.
Do I need an account to try it?: No. Paste a PDF link into the reader and read the result.
What does it cost to keep using it?: Making an account starts a 14-day trial of the full product with no card asked for. After that you pick how often you pay: $10 billed monthly, $60 billed once a year, or $108 billed once every 3 years. Every one of them is the same whole product.
What happens to PDFs I already saved if I stop paying?: You keep reading them. The account goes read-only: the readlist, the reader view, the extracted text and export all keep working. Saving new links and importing are what stop.

What it costs after the first one

Reading a link you paste here costs nothing. Keeping a library of them is a subscription: 14 days free, no card. After that $3/month.

If you never subscribe, nothing is charged and the account goes read-only, not dark — you keep reading every article you saved, and you can still export.

Google, Apple, or an email address. No card at any point in the trial.

Try it on a PDF you already have

Paste a link and read the extraction. No account required. Reading a PDF on a phone and turning one into an EPUB have their own pages.