Talk: RAG-as-a-service in production

Watch on YouTube (opens in a new tab)

My talk at the Intelligence Artificielle Aix-Marseille meetup #7 (in French, 48 minutes): Construire un RAG-as-a-service en production : architecture et REX. The slides are on SlideShare.

During my mission at ADEO (Leroy Merlin’s parent company), our team built a platform that lets any of the group’s 115,000 employees create a chatbot on their own documents in minutes, without technical skills. We built it because employees lose a lot of time hunting for information spread across many tools.

What I cover

  • Why we chose RAG over fine-tuning for company knowledge
  • Incremental ingestion pipelines with ZenML (Google Drive, GitHub, Confluence, GCS)
  • Serving with LlamaIndex workflows, guardrails, and per-chatbot configuration stored in a database and applied at runtime
  • What it takes to run in production: SSO, Vault, traceability, observability, and evaluation with Ragas
  • A live demo, plus Q&A on governance, access rights and how to measure value

At the time of the talk: 338 chatbots created by 161 people, about 47,000 synced documents and more than 1 million vectors.