Document recognition with local AI

Documents in. Data out.

docsniff reads invoices, delivery notes and receipts and returns validated, structured data. No cloud AI.

Muster Bürotechnik GmbH Beispielweg 12, 12345 Musterstadt
Rechnung
Nr. RE-2026-04187 Datum 14.10.2026
Toner-Kartusche schwarz, 2 x 89,90179,80
Wartungspauschale Q4240,00
Netto419,80 EUR
USt 19 %79,76 EUR
Gesamt499,56 EUR
re_2026_04187.pdf Done
Document type
invoice_de98 %
Supplier
Muster Bürotechnik GmbH97 %
Document number
RE-2026-0418799 %
Date
2026-10-1499 %
Net
419.8098 %
VAT
79.7697 %
Gross
499.5699 %

Totals and tax rate plausible

Sample documents. All names and details are fictitious.
  • Invoice DE
  • Factuur NL
  • Facture FR
  • Invoice GB / IE
  • Credit note
  • Delivery note
  • Receipt
  • PDF
  • JPG
  • PNG
  • TIFF
  • HEIC

From scan to record in six steps.

Every document runs through the same pipeline. Scroll along and see what happens to it.

  1. Capture

    Via REST API or browser upload, soon also via Hot Folder. docsniff spots duplicate documents by their SHA-256 hash.

  2. Text recognition

    If the PDF has a text layer, it is used. Scans and photos are read by Tesseract, word by word with a confidence score.

  3. Classification

    The local language model determines the document type: invoice from DE, NL, FR or GB, credit note, delivery note or receipt.

  4. Extraction

    Each document type has a fixed JSON schema. The model must return valid fields, free-text answers are not possible.

  5. Validation

    Calculation rules, country-specific tax rates and tolerances check every result. Uncertain documents go to the review view.

  6. Handover

    Your software retrieves the structured data via REST API, webhooks are planned. You then archive the documents in your own archiving system.

Built for everyday paperwork.

Two ways in, soon three

REST API for your software and upload in the browser. In preparation: a Hot Folder per client for scanners and file servers.

Not sure? Then a human checks.

Below the threshold, the document lands in the review view, with the original and the fields side by side.

Validation rules per client

Checks on or off, tolerances and tax rates per country.

  • DE19 / 7
  • NL21 / 9
  • FR20 / 10 / 5,5
  • GB20 / 5

Four languages

Interface in German, Dutch, French and English.

No document twice

Every file gets a SHA-256 fingerprint. If it arrives a second time, it is recognised instead of being booked again.

Your documents never go to a cloud AI.

Text recognition, language model and database run on the same infrastructure. No document is sent to OpenAI, Google or any other AI provider.

Language model on site

Ollama with a Qwen model on a dedicated GPU. No request leaves the server.

Clients strictly separated

Every record belongs to exactly one client. Without a client set, the database returns nothing at all.

Retention on your terms

You decide per client how long docsniff keeps documents. Audit-proof archiving remains with your archiving system.

Access only from your network

IP ranges allowed per client, plus roles for users and administrators.

Coming soon

docsniff On-Premise

The same software in your own data centre, on your hardware. Not available yet. Interested? Just mention it in your enquiry.

Frequently asked questions

Which files and documents does docsniff understand?

PDF, JPG, PNG, TIFF and HEIC. It recognises invoices from Germany, the Netherlands, France, the United Kingdom and Ireland, as well as credit notes, delivery notes and receipts.

What about XRechnung and ZUGFeRD?

These German e-invoice formats already contain their data in machine-readable form and need no text recognition. docsniff is designed for all documents where that data is missing: paper, scans, photos and PDFs without XML.

Where do my documents go?

Only to the docsniff infrastructure. Text recognition and the language model run there, with no calls to external AI services.

What happens if the AI is unsure?

If the confidence is below your client threshold or a validation rule is triggered, the document goes to the review view. There you correct the fields with the original right beside them.

How do I connect docsniff to my ERP or DMS?

Your software retrieves the data as JSON via the REST API. Webhooks and a Hot Folder for scanners and file servers are planned. We will work out together which route suits your software.

What availability applies?

The standard SLA provides for 95 % availability. A higher SLA can be agreed on request.

Does docsniff replace my archive?

No. docsniff is not an archiving solution under GoBD, the German rules for compliant bookkeeping and archiving. For audit-proof retention of your documents you still need your own archiving or DMS system.

Is docsniff also available on-premise?

Not yet. A version for operation in your own data centre is planned. If you are interested, feel free to mention it in your enquiry.

Who is my contact?

Sales and consulting are handled by INOSYS GmbH in Würzburg. docsniff is operated by i can GmbH, also based in Würzburg.

See docsniff with your own documents.

Tell us which documents you deal with. Our sales partner INOSYS GmbH will get back to you with a date for the demo.

+49 931 35286-0 info@inosys.de

INOSYS GmbH, Äußere Aumühlstraße 13, 97076 Würzburg

For example, which software you use today.