On-premise OCR API with word boxes and confidence scores
A cloud-style OCR API that runs on your own server, for sensitive documents and for AI agents that should spend fewer tokens on page images: send a PDF or an image, get back text, a box and a confidence for every word. No call to an outside AI service.
Production or your own server? Ask for an offer
OCR with word bounding boxes and table cells, down to one character
Five distances on one ugly invoice: skewed, speckled, stamped. From one character to the whole page, every tile shows the box you get.
Character A zero, not the letter O. Box measured on the image.
Word One word,
Northlight. Box measured on the image.Line Three lines, each with its box and a confidence of 0.99, 0.99 and 0.99.
Table Every cell of the first four rows with its own box, each 0.99 sure or more.
Page Every line and cell of the page, each with its box.
Watch a whole page being read, step by step: gallery of test pages.
Versioned OCR pipelines for scans, rotated pages, tables and hard pages
Every pipeline is a fixed chain of OCR engines, chosen after measuring many models on the same test pages.
OCR for everyday office scans
Typical office scans, with text and place for every line or word.
Rotated page OCR: auto-rotate at 90, 180 and 270 degrees
Found at 90, 180 or 270 degrees and read upright.
Table OCR: a box for every cell
Every cell gets its own box, and a grid that could not be built is marked as failed.
Hardest pages: language model plus character engine
A model reads the whole page; a second OCR engine flags the words where they disagree.
An improvement is a new version; the old one stays. All pipelines at a glance · Every reading mode and its terms.
OCR accuracy and speed: pages per hour on hard scans
No single OCR reads every page best, so the reading modes differ. On the hardest test pages you see what each finds, how fast, and which words it flags, so you can pick the one that fits your documents.
One line of a hotel invoice, three OCR engines, three different readings.
- Image (cropped)
- OCR engine A NOĆENJE/DORUČAKEvery letter as printed.
- OCR engine B NOČENJE/DORUČAKČ instead of Ć: the wrong mark.
- OCR engine C NOCENJE/DORUČAKC instead of Ć: the mark is lost.
Self-hosted OCR API: sensitive documents stay in your network
Moving from a cloud OCR service? The four questions buyers ask first, and how each is answered here.
On-premise OCR: documents stay in your own network
Your documents never leave your own network: on-premise OCR with no transfer to a cloud service, so sensitive documents stay under your own control. You download the package, check its SHA-256 and start it on your own server; after the install it runs offline.
Your own server →
Reproducible OCR with pinned pipeline versions
A pipeline is named and versioned, and a pinned version never changes under you. Fast reading returned identical bytes at every batch size.
Pipelines and versions →
Confidence score per word for digits that must be right
Every word comes with a box and a confidence. On the hardest test pages, the language-model pipeline returned nearly every number exactly as printed, and every wrong value carried a flag.
How a second reader catches it →
Batch OCR API: 1,000 pages in one request
Send up to 1,000 pages in one request, wait for the job, then fetch the result in one call.
API guide →
OCR to JSON, text, Markdown, hOCR and searchable PDF
Send PDF, PNG, JPEG or TIFF. Get the format your next tool reads.
- PNG
- JPEG
- TIFF
- JSON
- TXT
- MD
- hOCR
OCR to JSON
Every word with place and confidence.
JSON →
OCR to plain text
The words in reading order, nothing else.
Text →
OCR to Markdown
One section per page, safe to read.
Markdown →
hOCR output
HTML with a box for every line or word.
hOCR →
Searchable PDF from scans
Your original with a text layer.
Searchable PDF →
On-premise OCR software: where it runs, licence, support
- Where
- On your own Linux server with an NVIDIA card: download the package and run the install script
- Your files
- Your documents are processed on your machines and stay in your house
- Licence
- Free for private use and tests under the EULA that comes with the download; productive use needs a contract
- On request
- Operation for third parties, support and maintenance: price on request
Questions from a purchasing team: Answers for procurement · What a company pays and what each way covers: Pricing
Managed OCR server in Germany, custom OCR pipeline, language model on OCR
Services for larger customers. Price on request.
Managed on-premise OCR server in Germany
A dedicated server in Germany, operated and maintained for you, with only your pages on it.
Managed OCR server →
Your own pipeline
A pipeline made for your documents, named and versioned as name@version.
A language model on top
Parsing, interpretation and archive search that work on grounded OCR.
Try the OCR API free: gallery, download, first call
Faxes, stamps, coffee stains: see how your scans read.
Watch a page being read.
The gallery shows a public test page, word by word.
Download.
Install the release on your own hardware.
Send pages.
Three steps: send, wait, fetch.
The whole API, with every call: API guide. Machines: agent guide, openapi.json, llms.txt.
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.
