RAG (Retrieval-Augmented Generation) was proposed by Meta AI in 2020, compensating for the limitations of a model's parametric knowledge by dynamically retrieving external knowledge bases during LLM inference.