RAG-Sequence

RAG-Sequence uses a single retrieved document to guide the generation of an entire output sequence for maximum topical consistency.

Lewis et al. (2020) introduced RAG-Sequence as a specialized Retrieval-Augmented Generation formulation that treats the retrieved document as a latent variable for the complete response. Unlike RAG-Token (which switches documents per word), this model retrieves a set of top-k documents and marginalizes their probabilities to produce a cohesive narrative. It excels in knowledge-intensive tasks like Natural Questions or Jeopardy! where maintaining a specific context across multiple sentences is critical for factual accuracy.

https://arxiv.org/abs/2005.11401