Document archive that auditors accept on the first attempt

Documents are stored in unchangeable form with a SHA-256 hash checksum, a time stamp and a complete audit trail. Changing the original is technically impossible: the system only allows a new version, and the history stays complete. Retention periods come pre-configured per document type, with 10 years for booking documents, 6 years for business letters and 30 years for patents and personnel files. Once a period expires, the system automatically suggests those documents for deletion. You confirm in the four-eyes principle, and the deletion is documented with an audit entry that includes the hash of the deleted file. The auditor export is ready at the press of a button, with all documents, audit trails and hash verification for machine-readability under GoBD.

Every uploaded PDF, image or scan runs through OCR (Tesseract, EU on-prem) and is indexed for full-text search. You search for 'delivery note 2024 Müller GmbH' and the system finds every matching document, even in mail attachments, in scanned contracts and in handwritten notes, although the hit rate for handwriting is limited. The search supports wildcards, Boolean operators, phrase search and searching in metadata such as author, creation date, file and keywords. Hits are highlighted right in the document, so you jump straight to the right page. OCR quality for standard scans is over 98 percent, and where scan quality is poor the AI helper suggests corrections. Multilingual OCR for German, English, French and Italian is included by default.

Every change to a document creates a new version, while the original stays archived and unchangeable. The comparison view shows two versions side by side with the changes highlighted, as Word diff, Excel diff or PDF diff. For each version you see who changed what and when. Rolling back to an older version takes a single click: the old version becomes the new one, and the versions in between stay untouched in the history. This matters especially for contracts, because you can prove that the signed version is unchanged, with the hash of the original file documented in the audit log. For eIDAS-signed documents the signature is re-validated on every version change.

Organise your documents in files, for example per customer, per project, per employee or per supplier. File structures work as templates: when you create a new customer file, all the standard folders appear automatically, from contracts and correspondence to invoices, delivery notes and complaints. Permissions apply per file or per folder, so a sales rep sees only their own customers, HR sees only employee files, and external partners get time-limited read-only access to one specific file. On upload, the AI helper automatically suggests the right file and folder based on content, sender and document type. Multi-tenant isolation via Postgres Row-Level Security makes sure no data leaks between tenants.

With the Outlook and Gmail plugin, one click archives a mail including its attachments straight into the right file, and the AI helper suggests that file based on sender, subject and mail thread. Word, Excel and PowerPoint files open directly from DARION-AI; when you save them back, a new version is created automatically, and several people can co-edit at the same time via Microsoft 365 or Google Workspace. Both Microsoft 365 and Google Workspace connect via OAuth, so no third party gets to copy your documents. The web clipper, a browser extension, archives web pages as PDF including the source URL and date, which matters for research documentation and compliance records.

Upload a document and the AI helper (Qwen3, self-hosted in the EU) classifies it automatically by document type, whether invoice, contract, delivery note, letter or quote, extracts metadata such as date, sender, amount and document number, and suggests keywords. The hit rate for standard documents is typically over 90 percent, and the AI helper learns from your corrections. The contract clause analysis points you to critical passages such as notice periods, liability limitations, price adjustment clauses and confidentiality agreements. For sensitive files you can switch off AI processing per file; in that case only OCR runs, without any AI models. Sensitive data never leaves the EU, and the model is trained exclusively on your anonymised data.