100% Sovereign ChatBot/RAG
REXAI in production
2025-12-11 | 11:55 AM - 12:20 PM | Amphitheater Gaston Berger I Level -2
Information
In early 2025, we were contacted by a client in the Renewable Energy sector who asked us to sign an NDA before even starting discussions. We quickly understood that the security of data, their processes, and IP was a fundamental element of their expertise.
They wanted to create a ChatBot/RAG that would allow them to exploit a huge pool of unstructured data (doc, docx, pdf, images, xls, etc.) mainly to improve their R&D processes (subjects around technological monitoring and patents), and support (instant questions/answers)
We (Accure) are a consulting firm specializing in Data and Cloud and we initially thought of so-called "classic" solutions from a hyperscaler such as Azure, AWS or Google which would allow us to achieve our objectives quickly by putting together the building blocks we would need. However, the client very quickly stopped us by demanding 100% sovereign, 100% French, no Cloud provider or services that could fall under the Cloud-Act or Patriot-Act as their know-how is critical.
We took up the challenge and put v1 into production at the beginning of July 2025!
This feedback will therefore focus on the following topics:
- Governance / Sovereignty: in 2025 how to make "made in France hosted in France" while remaining at the cutting edge of technology
- AI stack in production: what was our thinking to provide an AI that does "as well as the big ones" while remaining sovereign.
- DevOps / DataOps: implementation of CI/CD to deliver the platform, the code. Lifecycle of data and AI outputs for continuous improvement of the solution
- FinOps: what were our thoughts on costs and how to contain them while still providing the service?
Technical stack:
- PAAS Hosting (Clever Cloud): Docker + S3
- StreamLit / Python
- Postgres 17 + pgVector
- Heptapod (Gitlab open source)
- Mistral AI (Embed + AI) via API
- Airflow 3
