How is an AI supposed to know what only your business knows?
A language model writes any general advice text. What it cannot do: name your project from last autumn, your price range, your region, your certificates. Exactly from those arises a text that works as a source.
What the area does
Read in documents — checked before they are read
PDF, DOCX, XLSX, TXT, CSV and Markdown. Every file passes a virus scan before the parser; if the scanner does not answer unambiguously, the file is not parsed. The parser itself has no database access — it sees the file and nothing else.
Facts with source, confirmation and expiry
Every number, every price, every certificate exists as its own object: with source, trust level, confirmation date and end of validity. Expired entries drop out of new texts instead of propagating themselves.
Only a human can confirm — and only with a location
A statement only counts as a fact once somebody has confirmed it, and confirming is only possible with a concrete section from a processed document as evidence. The decision is logged with the user account.
The quality gate checks against the fact store
The briefing feeds the generator exclusively confirmed facts. Numbers without cover are recognised by the check and named in the report — before approval, not after.
How you can tell
- ✓In the editor, every paragraph states what it relies on. What is not evidenced you see before approval.
- ✓Every fact carries the location in the original document, not just the file name.
- ✓Deleted sources disappear immediately from search and from new confirmations.
- ✓Lexical search over your knowledge always works. Semantic search is off by default; the interface shows per project which of the three conditions is missing — provider, your approval, or a complete index — instead of falling back silently.
Where it stops
- ×Without your own knowledge this software does not help either. Where there is nothing of your own to say, no citable text arises either.
- ×No fact arises automatically. A human must confirm, and without evidence it is not possible.
- ×Semantic search requires an external provider and your explicit approval — it is not switched on tacitly.
- ×The quality gate recognises unevidenced statements. It cannot guarantee that an evidenced text is also well written.
Frequent questions
Does the software just write texts with ChatGPT?
The difference lies before the writing. First your operational knowledge is captured: services, brands, regions, reference projects, price ranges, certificates — each with source and confirmation. Only from those does the text arise, and every number in it must point to a confirmed fact. Statements without evidence block approval.
What happens to a price that no longer applies?
It drops out. Every fact carries an end of validity; once it has passed, it is no longer fed to new texts. That is the reason for the date: without it an outdated price propagates through every new article, and nobody knows any more where it came from.
Are my documents safe from other customers?
The separation sits in the database, not in the application. Every query runs with the tenant context set; if it is missing, every tenant-bound query returns zero rows instead of all rows. This is verified with a dedicated test gate that queries deliberately without a filter — it checks whether the database holds when the application forgets to filter.