Banque des Territoires, Polyconseil, and Hugging Face Deploy Sovereign RAG Solution for EduRénov Program

TL;DR

The Banque des Territoires, Polyconseil, and Hugging Face have delivered a sovereign, open‑source Retrieval‑Augmented Generation (RAG) system that automates email responses for the EduRénov program, enabling faster, source‑cited support for the ecological renovation of 10,000 French public schools while keeping data and compute under French control.

Executive summary

The first phase of a joint project between Banque des Territoires (BdT, part of the Caisse des Dépôts et Consignations), consulting firm Polyconseil, and AI platform Hugging Face has produced a prototype RAG application. The tool is designed to streamline the email‑based interactions that underpin the EduRénov program, France’s flagship initiative to finance and coordinate the energy‑efficient renovation of 10 000 public school facilities. By leveraging open‑source models and a French‑certified cloud provider, the solution satisfies strict data‑sovereignty requirements while demonstrating the scalability of generative AI for public‑sector workflows.

Why RAG fits the EduRénov mission

EduRénov aims to support 10 000 renovation projects—about 20 % of France’s school infrastructure—and achieve a 40 % reduction in energy consumption within five years, backed by €2 bn in loans and €50 m for preparatory engineering. The program’s success depends on extensive, repetitive email exchanges between BdT experts, CDC officials, and local authorities. These exchanges draw on a large shared documentation base, making a Retrieval‑Augmented Generation system an ideal way to surface relevant information quickly and consistently.

"EduRénov has found its projects and cruising speed; now we will enhance the relationship quality with local authorities while seeking many new projects," said program director Nicolas Turcat.

Public entities have been cautious about generative AI because of data‑sensitivity and sovereignty concerns. Open‑source LLM ecosystems now match proprietary performance, allowing the CDC to pilot a data‑transformation project that guarantees sovereign compute and model usage.

Ensuring performance and sovereignty at scale

Early experiments with Hugging Face’s open‑source stack (Transformers, Sentence‑Transformers, Tokenizers, Text Generation Inference) confirmed the feasibility of a RAG approach. To meet French public‑sector security standards, the consortium chose:

  • Polyconseil – an Agile‑focused digital‑innovation consultancy with certified on‑premise and cloud deployment expertise (AWS Advanced Tier Services partner, GCP Cloud Architect, Kubernetes CKA). Polyconseil provides the full stack of data scientists, engineers, DevOps, UI/UX designers, and product managers.
  • Hugging Face – the global hub for over 750 000 models and 175 000 datasets, offering the Text Generation Inference (TGI) service for high‑performance, containerised LLM serving.
  • NumSpot – a French cloud provider with SecNumCloud certification, built on Outscale’s IaaS (Dassault Systèmes), ensuring data residency and compliance.

The solution therefore runs open‑source models on sovereign French infrastructure, eliminating reliance on foreign commercial APIs.

Modular architecture for a dynamic sector

The application lets BdT agents paste an incoming prospect email, automatically generates a draft response sourced from EduRénov documentation, and allows the agent to edit and approve the reply before sending. The workflow follows these steps:

  1. Client email received via existing channels.
  2. Email forwarded to the new user interface.
  3. Orchestrator builds a retrieval query.
  4. Retriever fetches relevant passages from a vector database (embedding‑based search).
  5. Orchestrator creates a prompt that includes the retrieved context and source citations.
  6. Reader module (LLM) generates a suggested reply.
  7. Agent reviews, edits, and validates the suggestion, providing RLHF‑style feedback.
  8. Validated reply is sent through the messaging system.
  9. Recipient receives the answer with explicit source references.
  10. Public repository of used documents is available for verification.

Four subsystems support this pipeline:

  • User interface – ingestion of documents, creation of fine‑tuning and RLHF datasets (green).
  • Messaging layer – integration with existing email systems (black).
  • RAG core – retrieval and generation components (purple).
  • Feedback & training store – pipeline logs, fine‑tuning data, and RLHF feedback (red).

Key success factors

  • Model‑agnostic modularity – LLM, embedding model, and retrieval method can be swapped via configuration, enabling rapid experimentation as the state‑of‑the‑art evolves.
  • Modular monolith with hexagonal architecture – isolates business logic while delegating GPU‑intensive LLM inference to an external TGI service.
  • Open‑source French models – currently uses Mistral‑7B‑Instruct‑v0.3; other candidates such as CroissantLLM (a collaboration between CentraleSupélec’s MICS lab and Illuin Technology) are under evaluation for French‑language ethics and performance.
  • Agile ScrumBan team – a single cross‑functional squad conducts weekly AI‑breakthrough briefings led by a Hugging Face Expert Support specialist, ensuring knowledge transfer to BdT’s data team.
  • Robust evaluation – early MVP includes qualitative and quantitative notebooks using the RAGAS library, providing a reproducible benchmark for future improvements.

"Based on this pilot, this approach opens a new pathway: it is likely how public policies will be deployed in the territories in the future, together with the necessary financing for the country's ecological and energy transformation," said Hakim Lahlou, Innovation and Strategy Director at BdT.

Implications for public‑sector AI

The case study demonstrates that sovereign, open‑source generative AI can be deployed at scale for mission‑critical public programs without compromising data residency. By combining a modular RAG pipeline, French‑certified cloud infrastructure, and locally trained models, the consortium provides a blueprint for other government initiatives that require secure, high‑throughput document‑centric automation.

If you face sovereignty challenges in your AI projects, contact Polyconseil at alivia@polyconseil.fr. For Hugging Face Expert Support inquiries, reach out via the Hugging Face sales contact page.

Sources