news

A simple RAG baseline matched the complicated pipelines

Stanford researchers report that keeping retrieved passages in their original document order matches or beats multi-stage retrieval pipelines.

research

What changed

In Stronger Baselines for Retrieval-Augmented Generation with Long-Context Language Models, revised January 2026, Alex Laitenberger, Christopher D. Manning and Nelson F. Liu report that a straightforward retrieve-then-read baseline which preserves passages in their original document order matches or outperforms the multi-stage pipelines ReadAgent and RAPTOR across several long-context question-answering benchmarks.

Why you should care

Before you add a reranker, a summariser and a tree index to a client's pipeline, measure the simple version, because it may already be your ceiling.

Read the source