Grounding

A language model knows a lot about the world, but nothing about your tariffs, deadlines and processes. When a customer asks about them, the answer has to come from your sources. That is what grounding does: it anchors the answer in verified knowledge.

#How RAG works

Retrieval-augmented generation means exactly that: generation extended by retrieval. The model does not write from memory, but from material laid out for this specific question.

1. Understand the question. The request is recognised and prepared for search.
2. Retrieve. A semantic search finds the matching passages in the knowledge base, even when different words are used.
3. Answer. The model phrases the answer only from the passages it found.
4. Cite. The answer names its source, so agents and customers can check it.

Univerx Knowledge

Question

Where is our data processed?

In EU regions of your choice. Hosting in Germany is planned. The sub-processor list and the processing-agreement documents are available on request.

SourceTrust Center · Data protection

#What drives quality

Grounding is only as good as the sources it draws on. The most important levers therefore sit not in the model, but in knowledge management.

• Freshness. Outdated articles produce outdated answers. Every piece of content needs an owner and a review date.
• Clarity. Two contradictory articles on the same topic lead to contradictory answers.
• Structure. Short sections with clear headings are found more reliably than long documents.
• Permissions. Internal notes do not belong in answers to customers.

#When nothing is found

Behaviour at the gaps matters just as much. If the search finds no matching source, the model must not improvise. The right answer is then: there is no verified information on this, and the request goes to a person.

#Grounding in Univerx

Univerx Knowledge is one knowledge base for people, bots and self-service, with semantic search and answers based on your sources. AI agents and Univerx Assist draw on the same base. Why that keeps AI honest is the topic of Grounded answers.