Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks

Abstract summary

Introduces Retrieval-Augmented Generation (RAG): an architecture that combines a pretrained sequence-to-sequence model (BART) with a non-parametric memory (a Dense Passage Retrieval index over Wikipedia). Demonstrates strong performance on knowledge-intensive NLP tasks while providing transparency about which documents informed each generation. Establishes the design pattern of retrieving documents before generating.

Our summary in our own words — see the canonical source links below for the original abstract.

Why we cite this paper

RAG is the architecture behind essentially every enterprise LLM deployment in 2026. The retrieval layer — typically using embedding models plus vector databases (Pinecone, Weaviate, Qdrant, Milvus) — is one of the densest engineering-acceleration surfaces we track at /sector/database and /trend/ai-native-databases-2026.

Key findings

• 1Combining parametric (model weights) and non-parametric (retrieved documents) memory outperforms either alone on knowledge-intensive tasks.
• 2Retrieval enables citation transparency — the model's outputs can be traced to specific source documents.
• 3RAG addresses three core LLM limitations: knowledge cutoff dates, hallucination on out-of-distribution facts, and inability to cite sources.
• 4The design pattern became the foundation of the modern vector-DB-backed LLM application stack.

Canonical sources

Related glossary terms

Frequently Asked Questions

Other research papers