RAG: Retrieval-Augmented Generation

A model has never seen your company's private documents, but your application can search those documents for passages related to a question. It then adds those passages to the request, and the model answers from them. No one retrains the model for this. This technique is called retrieval-augmented generation (RAG).

Take a support bot for an online store. A customer asks how many days she has to return an item. Without retrieval, a model might answer 90 days, because that number sounds likely, while the store's policy says 30 days