Trusted by Design: Sovereign, Permission-Aware, and Secure

Grounded generation means: answers are generated exclusively from authorized enterprise sources – with traceable source citations down to passage level.

A user asks a question in natural language. The system finds matching passages in the enterprise index. Only passages the user is authorized to access are passed to the language model as context. 

The model generates the answer exclusively on this basis. The answer includes citations down to document and passage level. Two employees, the same question – two answers, each precisely within their own permission scope.

The difference from an ungrounded answer based on general model knowledge: such an answer may sound plausible, but it is not backed by evidence. Grounded generation makes every statement verifiable, authorized and auditable – and significantly reduces the risk of hallucinations.

1 – Data Layer. Sources, connectors, documents, metadata, permissions and vectors reside in one controlled layer. Access rights are applied through early binding during indexing – including permissions inherited from folders, groups and roles. There is no intermediate state in which unauthorized hits exist and “merely” have to be hidden.

2 – Permission-aware Passages. Permission checks apply at passage level: users only see context they are allowed to see – in hits, filters, autocomplete, previews, the LLM context and agent steps. When sources are combined, each passage is checked against its own source system; one source cannot override the permissions of another.

3 – Hybrid Retrieval. Full-text, linguistic, metadata, semantic and vector search are combined – for precise results even with specialist terminology, synonyms and multilingual content.

4 – Reranking. The most relevant passages are prioritized from the hits before they are passed to the language model – the quality of the context determines the quality of the answer.

5 – Grounded Generation. The language model answers exclusively on the basis of verified passages – with citations down to the passage.

6 – Eval & Audit. Groundedness, citation quality and retrieval accuracy are evaluated and models are compared; logs, user feedback and the audit trail make every answer traceable – including the steps taken by AI agents. This provides the accountability that internal audit, data protection and information security require.

Conclusion: Not “just another RAG”, but a controlled architecture: sovereign, permission-checked, evidence-based and auditable – and therefore more production-ready than generic assistant approaches.