We're Doing RAG All Wrong—and How We Can Do So Much Better

This talk challenges the conventional approach to Retrieval Augmented Generation (RAG), which typically involves embedding user queries, performing nearest neighbor searches on chunked text via vector databases, and fitting results into prompts. The speaker argues that this method is suboptimal and misses opportunities to maximize relevant information within the context window. The session will delve into why the current RAG strategy is limiting and explore smarter techniques for optimizing context, enhancing retrieval, and unlocking the full potential of RAG systems. The speaker, Simba Khadder, Founder & CEO of Featureform, draws on his experience with recommender systems and ML infrastructure.