While it has recently become widely accessible to develop a Proof-Of-Concept for Retrieval Augmented Generation (RAG) using OpenAI and one of the various open-source contribution on the topic, transitioning to a production-ready pipeline presents its own set of challenges. In this talk, Noé will share great practices for building a RAG product based on his experience developing Hikari, a bot utilized by hundreds of individuals within the Theodo Group, which continuously ingests documents from the group, such as those from Notion or HubSpot. These practices include (but not limited to) using a Directed Acyclic Graph (DAG) for continuous document ingestion (e.g., with Airflow), iterating on prompts, chunks, models, and more (e.g., with DVC), as well as understanding when, why, and how to switch to open-source models.

How to build apps on AWS using on demand serverless services and the newly released Amazon Bedrock.