France's CDC bets on open-source RAG to renovate 10,000 schools

France’s Banque des Territoires is betting that generative AI can accelerate one of the country’s most ambitious environmental programs. The institution, part of the Caisse des Dépôts et Consignations (CDC) group, has just completed the first phase of a sovereign AI tool built to support EduRénov, a program targeting the energy renovation of 10,000 public school buildings — nurseries through universities. That’s 20% of the national infrastructure pool, backed by 2 billion euros in loans and 50 million euros for preparatory engineering.

The problem was straightforward: program experts were drowning in repetitive email correspondence with local authorities, all of which required answers grounded in a massive shared documentation base. A Retrieval Augmented Generation (RAG) system — where a language model pulls from internal documents before generating responses — was the obvious fit. But for a public institution handling sensitive data, off-the-shelf commercial APIs were off the table. As program director Nicolas Turcat put it, “EduRénov has found its projects and cruising speed; now we will enhance the relationship quality with local authorities while seeking many new projects.”

Enter Polyconseil and Hugging Face. The CDC had already experimented with Hugging Face’s open-source stack — Text Generation Inference, Transformers, Sentence Transformers — and validated that a RAG approach could work. The consortium pairing Polyconseil’s agile development expertise with Hugging Face’s deployment knowledge was selected specifically to keep everything sovereign: compute, models, and data all stay within French control.

This matters beyond France. Public sector AI adoption has lagged precisely because the default path — OpenAI or Anthropic APIs — raises sovereignty alarms in Europe. The CDC’s project is a concrete test of whether open-source models have closed the performance gap enough to handle real operational workloads. Nearly 2,000 projects signed in the first year suggests the demand is there. Whether the tool can scale to handle thousands more without degrading response quality is the question the next phase will answer. If it works, expect other European public institutions to copy the playbook rather than wait for regulatory clarity that may never come.