Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks

Type: paper Tier: 2 (Foundational Paper) Author(s): Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, Douwe Kiela Date: 2020 (submitted 2020-05-22; NeurIPS 2020) URL: https://arxiv.org/abs/2005.11401 Accessed: 2026-06-01

Why This Source Matters

This is the paper that named RAG and is the primary source for the architecture behind Project 04. It frames the exact problem the PDF research assistant solves: a language model that "stores factual knowledge in its parameters" cannot precisely access or update that knowledge, and "providing provenance for [its] decisions and updating [its] world knowledge remain open research problems." RAG's answer — couple a generator with a retrieved, non-parametric knowledge store — is the design every production document-QA system still uses. It grounds the lesson's motivation, the parametric-vs-non-parametric mental model, and the "answers must be grounded and attributable" claim.

Key Claims

The problem with parametric-only models

• LLMs "store factual knowledge in their parameters" but "their ability to access and precisely manipulate knowledge is still limited, and hence on knowledge-intensive tasks, their performance lags behind task-specific architectures."
• Two open problems with parametric-only memory: "providing provenance for their decisions and updating their world knowledge." (You cannot cite why a parametric model said something, and you cannot edit a fact without retraining.)

What RAG is

• RAG models "combine pre-trained parametric and non-parametric memory for language generation."
• Parametric memory: a pre-trained seq2seq model (BART) — the generator's learned weights.
• Non-parametric memory: "a dense vector index of Wikipedia, accessed with a pre-trained neural retriever" (DPR). This is the swappable, updatable knowledge store — in Project 04, your indexed PDF chunks.
• The retriever finds relevant passages; the generator conditions on the query and those passages to produce the answer.

Two formulations

• RAG-Sequence: "conditions on the same retrieved passages across the whole generated sequence."
• RAG-Token: "can use different passages per token."
• (Project 04 uses the practical modern form: retrieve top-k chunks once, stuff them into the prompt, generate one grounded answer — closest in spirit to RAG-Sequence.)

The payoff

• Sets "the state-of-the-art on three open domain QA tasks, outperforming parametric seq2seq models and task-specific retrieve-and-extract architectures."
• "RAG models generate more specific, diverse and factual language than a state-of-the-art parametric-only seq2seq baseline." Grounding in retrieved text reduces hallucination and makes answers attributable.

Relevant To

• concepts: [retrieval-augmented-generation, rag, parametric-vs-non-parametric-memory, grounding, provenance, hallucination, knowledge-intensive-qa]
• projects: [04-pdf-research-assistant, 05-personal-memory-system, 07-ai-evaluation-framework]

Notes

• The modern "RAG" most engineers build is a simplification of this paper: retrieve top-k chunks from a vector DB (Project 03), concatenate them into the prompt as context, and ask the LLM to answer only from that context with citations. The end-to-end differentiable training in the paper is usually dropped; the retrieve-then-generate skeleton is kept.
• Why it beats a bigger model / longer context: non-parametric memory is updatable (re-index, don't retrain) and attributable (you know which passage produced the answer). Those are the two open problems the abstract names.
• The retriever here is exactly the bi-encoder + dense vector index pattern from sources/papers/sentence-bert.md and sources/official-docs/chromadb.md. RAG = Project 03's retrieval feeding an LLM.