The original Retrieval-Augmented Generation (RAG) framework, introduced by Lewis et al. (2020), is an end-to-end model that combines a dense passage retriever (DPR) with a pretrained seq2seq generator (BART). Given an input query, the retriever fetches the top-$k$ passages from a non-parametric memory (a Wikipedia index), and the generator conditions on the query together with the retrieved passages to produce the output sequence. The framework defines two variants: RAG-Sequence, which uses the same retrieved document for the entire generated sequence, and RAG-Token, which can attend to different retrieved documents per generated token, marginalizing over passages at each step. Both retriever and generator are jointly fine-tuned, making this the canonical reference cited by modern RAG systems — including this paper — for combining dense retrieval with sequence generation.