RAG is a term used to describe a technique to enhance the accuracy and reliability of generative artificial intelligence (AI) programs with facts fetched from external sources, according to an article on Nvidia.
Deeper Dive
Generative AI models such as ChatGPT can create original content based on inputs from a user, like generating a letter or recipe based on a short text request. But it has flaws: It sometimes presents false information when it doesn’t have the answer; can present old/outdated information; or create inaccurate responses. Why? Because it can only pull information from the dataset it was trained on.
RAG helps improve the responses by redirecting the large language model to retrieve relevant information from authoritative sources determined by a programmer. For example, it could be directed to search through footnotes in a research paper, or a new medical index. Businesses can use RAG to combine existing large language models with internal company databases to help their employees and customers.
For more in-depth information on how RAG works and its history, see this blog from Nvidia.com and this description from Amazon.com.