The context
The client's legal and business teams work on a dense documentary corpus combining regulatory texts, internal notes and case law. Consumer tools hosted outside the EU were ruled out for obvious confidentiality and Cloud Act reasons. Leadership was looking for an assistant that could answer internal questions quickly, drawing on the organisation's actual corpus, without exposing a single line of that content to a US provider.
The need
Three non-negotiable requirements. First, no data could leave European territory, neither at inference, nor in the logs, nor in the vector layer. Second, every answer had to be traceable back to a specific source passage, a condition of business use. Third, the solution had to be documented to the point where an internal team could operate and evolve it without any residual dependency on the firm.
What we delivered
A complete RAG assistant. Mistral model hosted on OVHcloud in a French region, Qdrant vector database self-hosted within the same perimeter, retrieval and synthesis steps orchestrated in LangGraph, access API in FastAPI. No outbound call to a non-EU provider, which closes Cloud Act exposure. The architecture document delivered sets out the data flow, the physical location of each component, and the mapping to AI Act and GDPR obligations applicable to the system. Operations documentation and training transferred to the client's teams.