Frequently Asked Questions

How is my allowance calculated?

Consumption per page depends on document type and proofreading. Standard uses 4 per page, Printed text (lightweight model) uses 2, and High accuracy uses 12. Automatic proofreading doubles it; choosing batch processing halves it. The free allowance is 120 per month (= 30 standard pages).

Are failed pages counted against my allowance?

No. If the system cannot complete a page (including abnormal AI output such as the same content repeated endlessly), that page does not consume your allowance. It is retried automatically a few times; if it still cannot complete, it is recorded as failed and not counted. Only pages that complete successfully are counted, and each job's detail screen shows the success/failure breakdown.

What materials work well? Can it read cursive (kuzushiji) or newspapers?

It works well on printed type and on relatively legible handwriting. It is not good at cursive Japanese scripts (hentaigana/grass script), complex multi-column layouts (e.g. newspapers), or badly degraded materials, and often will not reach usable accuracy on them. For transcribing cursive scripts we recommend a dedicated OCR (for example, "Miwo" from the Center for Open Data in the Humanities, CODH).

Famous classics or published books fail with a "recitation" error.

Gemini has a built-in "recitation" guard that stops generation when the output would nearly match a published work in its training data. A faithful transcription of a well-known text — say, a famous critical edition of Plato — can trigger it. This is not a problem with your material or settings, and it is not a bug in this service. Failed pages don't consume credits.

• Re-running (the "retry failed pages" button on the job page, or re-running the test) often succeeds — each attempt varies our internal counter-measures.
• Searchable PDF (coordinate OCR) and JSON mode are less prone to it than plain text mode, because their output format diverges from the original print.
• Pages that keep failing may be unavoidable for now due to Gemini's design; consider transcribing just those pages by hand.

The OCR result is wrong, or the layout is broken.

Accuracy varies greatly with the material and with your prompt. Try describing the language, typesetting, and the parts you want in your prompt, and use the Test feature on one page first. Note that, by the nature of AI, it can generate characters that do not exist in the original (hallucination), so always check results against the original. As long as processing completes, pages are counted even if the accuracy is insufficient — this is not a "failure" (recognition accuracy is not guaranteed).

How do I choose the document type (Standard / Printed / High accuracy)?

• Printed material (modern typeset books, journals) -> "Printed text". Lightweight model, half the credits.
• General handwritten or historical documents -> "Standard".
• Hard-to-read or important materials -> "High accuracy". Uses the top model, but is still weak on cursive scripts and complex layouts (see the question above).
• If you're not in a hurry, choose "batch processing" to halve consumption (results take longer).

Does turning on automatic proofreading increase consumption?

Yes. Turning on LLM proofreading doubles the credits per page (the AI is called twice — for OCR and for proofreading). It does not always improve accuracy, so use it when the test suggests it helps.

What happens if I use up my monthly allowance?

New jobs that month end as 'quota exceeded' and won't run. Your allowance resets on the 1st at midnight (JST).

How long can I download my results?

Results are kept for up to 30 days, then deleted automatically. Download links expire after 1 hour, but reopening the job detail screen issues a fresh link. Please save anything you need promptly.

How is my uploaded material handled?

Uploaded original files, processing images and result files are all deleted automatically 30 days after creation (uploads never turned into a job are deleted after 48 hours). If the searchable-PDF option is enabled, a searchable PDF (your original with a text layer added) is produced as a result file and deleted on the same schedule. Data older than 30 days cannot be recovered, so download what you need within the period.

Can I use it like an app on my phone?

Add it to your home screen and it launches full-screen, without the address bar. Chrome on Android offers this as “Install”. iPhone and iPad show no prompt, so open the Safari share menu and choose “Add to Home Screen”. It stays an ordinary web page either way, and it does not work offline. You can close the browser while a job runs — processing continues on the server.

I found a bug / I have a question.

This is a non-profit, free tool offered by volunteer researchers. Support is best-effort and individual replies are not guaranteed, but please report issues via the bug-report form (we improve things as we can): https://forms.gle/1NTz78iubG9bwrgQA