Research Papers

Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks

The foundational RAG paper — retrieval as a first-class input to generative models.

RAG established the pattern that most current answer engines still use: a dense retriever fronts a generative model, and the two are trained jointly.

Why this paper is worth re-reading

The core insight is that retrieval quality caps generation quality. If the retriever surfaces weak passages, the generator cannot rescue the answer.

That inverts the emphasis for source sites: passage-level quality matters more than page-level authority. A good passage on a mid-authority site can beat a mediocre passage on a strong one.