FAQ
OCR API FAQ: on-premise, data protection, accuracy, formats and licence
Before you trust an OCR with your documents you ask: where do my pages go, how wrong can it be, what does it commit me to? Whether you hold the scans, write the code, run the servers or sign the order, your question is probably here. One short answer each, and a link to the page that goes deeper.
OCR for scans and PDFs: accuracy, languages, try before install
Can I try OCR on a page before I install anything?
Yes. Open the gallery, pick a page and click a word to see its box and its confidence.
How accurate is the OCR on hard scans and fax pages?
On six hard pages with 135 target values, the character engine alone found 105, 5 of them silently wrong; the combination of a language model and a character engine found 131, 0 silently wrong (measured with the language-model pipeline). Your pages decide the rest, so a pipeline is measured on your own sample. Tables are measured too.
Which OCR languages and document types are supported?
German and English, measured on printed pages, scans, tables and turned pages. The answer holds text, boxes, confidence and layout; a language model on top can turn that into fields such as an invoice number or a total (how). There is no language setting: you send the page and get the text back. Another language matters to you? A pipeline can be built and measured for it. Documents go in as PDF, TIFF, PNG or JPEG. PDF to text and image to text say what to expect from a PDF or a photo.
OCR API for developers: batch requests, JSON output, reproducible results
How many pages can one OCR API request take?
Up to 1,000 pages and 100 MiB in one request, as PDF, TIFF, PNG or JPEG. No loop over single pages. You get back JSON, text, Markdown, hOCR or a searchable PDF. Output formats
Where do I start?
Read the API guide, which links the OpenAPI file and the full reference. Your own coding agent can do the setup from it.
Why do OCR results change between runs, and how do I get deterministic OCR?
A pipeline is named and versioned, and a version stays fixed. On 300 public test pages the CTC reading engine gave bit-identical results at every batch size. The language-model pipeline returns the same characters after every restart when it reads one page at a time (98 of 98 test pages); pick a fast reading pipeline when the output must be identical at any load. How the check works
How do I find wrongly read digits with a confidence score per word?
Every word comes back with a box and a confidence, so you can look at the digits first. Where two OCR engines read a word differently, the pipeline for the hardest pages flags it instead of guessing. How a character engine catches it
How does the OCR API report a page that failed?
The header `X-Pages-Failed` on the answer holds the number of failed pages; `pages_failed` in the job status names them.
Which OCR pipeline should I choose?
You choose the pipeline; the service reads with exactly that one. The pipelines at a glance list what each one delivers.
Self-hosted OCR: on-premise install, GPU sizing, restart, backup
Is there an on-premise OCR API for my own hardware, offline or in a private cloud?
Yes. The software is a download that runs on your own machine, on-premise or in a private cloud, with no call to an outside AI service. The internet is needed once to fetch the model weights; after that the service runs offline. Private use and tests need no application. Install it and see which GPU it needs.
How long does a pinned pipeline version stay available?
A version is never changed in place: a change is a new `@version`, and a job stays with the version it was sent to. On your own server the release you installed keeps running until you update it; the last three releases are kept, and one command switches back to the previous one. A fixed update promise is part of a contract. Update and roll back
Which GPU does an OCR server need for how many pages per hour?
Start from the pages per hour of each pipeline, measured on one GPU. Divide your volume by that figure. Your card and your pages will differ, so measure a sample on your own server. Which cards carry which pipeline
Does the OCR run on a CPU or on an NVIDIA GPU?
The package runs on an NVIDIA GPU, and the start check confirms the card.
What happens to running jobs when the OCR service restarts?
After a start the service continues the jobs that were unfinished: a page that was already read is not read or counted twice. Send the same `Idempotency-Key` again and you get the old job back. `ocrctl backup` writes a copy of the database and `ocrctl restore` puts it back (with the service stopped); keep the copy on another machine. API guide
Can the OCR server be run for me?
Yes: a server of your own, with only your pages on it, set up and kept up to date. OCR server for your private cloud
How was the OCR service code hardened and audited?
The code was searched for flaws, every finding was fixed, and all findings are listed by class. Data and security
OCR software licence, price, SLA and vendor
What does OCR software cost, and is there an SLA?
The package is free for private use and for tests. Productive use in a company, running it for others, and support need a contract. Pricing Support hours and response times go into the written offer, so you can put them next to your checklist.
Is there a vendor lock-in or a minimum term?
Your results come in standard formats, so you can take them anywhere. The download is free for private use and tests under its EULA; productive use is agreed in writing, before you pay. The hosted service has its own SaaS terms.
Who stands behind the software?
The company named in the imprint. Its servers are run by an operator certified to ISO/IEC 27001; proof of the certificate.
Data protection and data sovereignty: where sensitive documents stay
Where do my documents go: data residency and on-premise OCR
Your documents never leave your own network. With the download, on-premise OCR runs entirely on your server: no transfer to a cloud service, no copy at the vendor, and the data stays in the country and the building you choose, which settles data residency. Because only your own systems handle the documents, data protection stays in your own hands. This website is static and collects nothing through forms. Privacy notice
Is a data processing agreement (AVV, DPA) needed?
With the download no processor is involved, because your pages stay on your server. On a server run for you, a data processing agreement is concluded in writing; contact page.
Custom OCR pipeline and language model on OCR output
Can I get a custom OCR pipeline or an LLM on top of the OCR result?
Yes, as a project. Models can be combined to fit your documents (a pipeline made for your documents), and a language model can work on the result with every value pointing back to its place on the page (language model on top of OCR).
Why does LLM-based OCR hallucinate text?
Because it writes what is likely, not what is there. How a character engine catches it
Production use, high volume or your own server?
Private use and tests are free. For a company in production, a server of your own or a pipeline built on your pages, send the kind of page and the volume, and get a written offer, or a clear no when the fit is not there.
