Landmark Guide

12 Papers That Shaped Modern RAG

A beginner-friendly map of the ideas behind dense retrieval, retrieval-augmented generation, self-correction, and structured retrieval.

If you are trying to understand why RAG became the default way to make LLMs useful on real knowledge, this is the shortest serious reading path we know. We are not trying to cover the full history of information retrieval here. This page focuses on the modern LLM-era RAG stack: retrieve, ground, generate, critique, and improve the retrieval layer instead of pretending the model already knows everything.

Timeline

2020–2024

Core papers

Builder bonus

Start with the 3 must-read papers → Jump to full timeline →

If you only read 3 papers, start here

Most readers do not need every retrieval trick on day one. If you want the shortest path to modern RAG, start with the retrieval foundation, the canonical RAG paper, and the paper that shows how RAG became self-correcting.

Dense Passage Retrieval for Open-Domain Question Answering

The retrieval paper that made modern vector-style search central to open-domain QA and, later, RAG systems.

Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks

The canonical paper that gave RAG its name and its basic architecture.

Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection

The paper that shows how modern RAG moved from one-shot retrieval into retrieval decisions and self-correction.

The shortest useful history of RAG

Modern RAG did not begin as a chat feature. It emerged when dense retrieval got good enough, language models started learning to use external knowledge, and builders realized that grounding beats pretending the model already remembers everything.

Retrieval Gets Good Enough

Dense retrieval and retrieval-aware language models made it realistic to fetch useful passages instead of relying only on sparse keyword search or parametric memory.

RAG Becomes an Architecture

The field converged on retrieval plus generation as a repeatable systems pattern rather than a one-off QA trick.

Retrieval Quality Becomes the Product Problem

Once RAG worked, builders discovered that query formulation, black-box LLM integration, and context placement were the real bottlenecks.

Self-Correcting and Structured RAG

Modern RAG systems now critique retrieval quality, repair bad context, and use richer structures than flat chunks.

The 12 core papers

These are the papers we would keep if we had to explain modern RAG to a smart newcomer in one page. They are the papers that changed what people built next.

Retrieval Gets Good Enough

Dense Passage Retrieval for Open-Domain Question Answering

Vladimir Karpukhin et al.

2020Must readBeginner

Modern RAG only works because dense retrieval made semantic passage search strong enough to become the default starting point.

Why it matters

DPR is one of the clearest foundations for the vector retrieval layer that later became standard in RAG pipelines.

What changed after this

Open-domain QA, vector search, and passage retrieval stopped being a niche retrieval topic and became core infrastructure for knowledge-grounded LLM systems.

Who should read

Anyone who wants to understand the retrieval half of RAG before thinking about generation.

Read paper → Read next: REALM: Retrieval-Augmented Language Model Pre-Training →

Retrieval Gets Good Enough

REALM: Retrieval-Augmented Language Model Pre-Training

Kelvin Guu et al.

2020Must readIntermediate

A language model can be trained to use retrieval as part of how it learns, not just as a wrapper added later.

Why it matters

REALM helped establish the idea that external retrieval and language modeling belong in the same system, not in separate worlds.

What changed after this

Researchers took retrieval-aware pretraining more seriously, and RAG stopped looking like a hack layered on top of a frozen model.

Who should read

Readers who want the bridge from dense retrieval to retrieval-aware language models.

Read paper → Read next: Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks →

RAG Becomes an Architecture

Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks

Patrick Lewis et al.

2020Must readBeginner

Instead of forcing a model to memorize everything, retrieve relevant passages and generate with evidence in view.

Why it matters

This is the canonical RAG paper. It gave the field the name, the framing, and the basic architecture that much of industry still uses.

What changed after this

Grounded generation became a mainstream systems pattern for assistants, enterprise QA, and knowledge tools.

Who should read

Everyone. This is the single most important paper on the page.

Read paper → Read next: FiD: Leveraging Passage Retrieval with Generative Models for Open-Domain Question Answering →

RAG Becomes an Architecture

FiD: Leveraging Passage Retrieval with Generative Models for Open-Domain Question Answering

Gautier Izacard et al.

2021Must readIntermediate

How you fuse retrieved passages into generation matters as much as the retrieval step itself.

Why it matters

FiD became one of the cleanest architectural recipes for turning a pile of retrieved passages into stronger generative answers.

What changed after this

Multi-passage readers, passage fusion, and stronger generator-side grounding became part of the default RAG toolbox.

Who should read

Builders who want to understand why naive “retrieve then paste” is not the whole story.

Read paper → Read next: RETRO: Improving language models by retrieving from trillions of tokens →

RAG Becomes an Architecture

RETRO: Improving language models by retrieving from trillions of tokens

Sebastian Borgeaud et al.

2021Must readIntermediate

Retrieval is not just an app-layer patch; it can be a scaling strategy for the language model itself.

Why it matters

RETRO reframed retrieval as a serious alternative to “just train a much larger model,” which is a strategic idea that still matters.

What changed after this

Retrieval stopped looking like a concession for weaker models and started looking like a principled way to extend knowledge capacity.

Who should read

Anyone who wants the big-picture reason retrieval keeps coming back in frontier model design.

Read paper → Read next: Atlas: Few-shot Learning with Retrieval Augmented Language Models →

RAG Becomes an Architecture

Atlas: Few-shot Learning with Retrieval Augmented Language Models

Gautier Izacard et al.

2022Should knowIntermediate

Retrieval can boost generalization and few-shot behavior, not just factual lookup.

Why it matters

Atlas helped show that retrieval-augmented models can outperform much larger closed-book models on knowledge-heavy tasks.

What changed after this

RAG became easier to defend as a capability strategy, not only a freshness patch.

Who should read

Builders and researchers who want the strongest case for retrieval as a capability multiplier.

Read paper → Read next: HyDE: Precise Zero-Shot Dense Retrieval without Relevance Labels →

Retrieval Quality Becomes the Product Problem

HyDE: Precise Zero-Shot Dense Retrieval without Relevance Labels

Luyu Gao et al.

2022Must readBeginner

Sometimes retrieval gets better when the model first imagines the ideal answer and retrieves against that.

Why it matters

HyDE is one of the cleanest examples of LLMs improving the retrieval side itself, not just the generation step.

What changed after this

Query expansion, hypothetical documents, and LLM-assisted retrieval setup became practical builder patterns.

Who should read

Anyone who cares about fixing retrieval quality without expensive supervision.

Read paper → Read next: REPLUG: Retrieval-Augmented Black-Box Language Models →

Retrieval Quality Becomes the Product Problem

REPLUG: Retrieval-Augmented Black-Box Language Models

Hong-Wei Wu et al.

2023Must readIntermediate

You can improve a black-box LLM with retrieval even when you do not control the model weights.

Why it matters

REPLUG pushed RAG closer to the real builder environment: APIs, black-box models, and product teams that control the system but not the base model.

What changed after this

Plug-in retrieval layers, API-first RAG stacks, and model-agnostic grounding became easier to productize.

Who should read

Builders shipping on top of hosted models rather than training their own.

Read paper → Read next: Lost in the Middle: How Language Models Use Long Contexts →

Retrieval Quality Becomes the Product Problem

Lost in the Middle: How Language Models Use Long Contexts

Nelson F. Liu et al.

2023Must readBeginner

More context is not automatically better if the model uses the middle of the prompt badly.

Why it matters

This paper explains why long context does not eliminate the need for good retrieval, ranking, and context placement.

What changed after this

Context ordering, chunk selection, and retrieval precision became more important product concerns than simply buying longer context windows.

Who should read

Everyone building RAG in production. This is where many practical mistakes become obvious.

Read paper → Read next: Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection →

Self-Correcting and Structured RAG

Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection

Akari Asai et al.

2023Must readIntermediate

A better RAG system decides when to retrieve and critiques its own generations instead of blindly following one fixed pipeline.

Why it matters

Self-RAG is a strong marker of modern RAG maturing beyond one-shot retrieval into adaptive retrieval and self-critique.

What changed after this

Retrieve-or-not decisions, critique tokens, and reflective generation loops became central ideas in advanced RAG systems.

Who should read

Builders who want to understand the jump from plain RAG to adaptive RAG.

Read paper → Read next: Corrective Retrieval Augmented Generation →

Self-Correcting and Structured RAG

Corrective Retrieval Augmented Generation

Xisen Jin et al.

2024Must readIntermediate

When retrieval is weak, a strong RAG system should notice that and repair the evidence pipeline before answering.

Why it matters

CRAG makes retrieval quality itself a first-class object of correction instead of assuming the retriever already did its job.

What changed after this

Fallback retrieval, retrieval grading, and evidence repair became more standard parts of serious RAG stacks.

Who should read

Anyone who has seen a good generator ruined by bad retrieval and wants the clearest paper on the problem.

Read paper → Read next: From Local to Global: A Graph RAG Approach to Query-Focused Summarization →

Self-Correcting and Structured RAG

From Local to Global: A Graph RAG Approach to Query-Focused Summarization

Hao Xin et al.

2024Must readIntermediate

Flat chunk retrieval is not always enough; some questions need structure over entities, relations, and higher-level summaries.

Why it matters

This paper captured why GraphRAG became such a popular term: many enterprise and research questions need structure, not just nearest-neighbor chunk lookup.

What changed after this

Graph-based retrieval, hierarchical summarization, and corpus structure became much more central to advanced RAG design.

Who should read

Builders dealing with large knowledge bases, cross-document reasoning, or questions that need synthesis rather than one snippet.

Read paper → Read next: RAGAs: Reliable and General Evaluation for Retrieval-Augmented Generation →

5 builder bonus papers

The core 12 explain the arc of modern RAG. These five bonus papers explain embeddings, late interaction, evaluation, hierarchical retrieval, and reranking patterns that matter in production.

Builder Bonus

Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks

Nils Reimers et al.

2019Must readBeginner

Fast semantic search starts with strong sentence embeddings, and Sentence-BERT made that practical.

Why it matters

Sentence-BERT is one of the simplest starting points for understanding why embedding-based retrieval became so useful in practice.

What changed after this

Semantic search and embedding-first retrieval became much easier to ship, which later fed directly into RAG tooling.

Who should read

Builders who want the shortest path to understanding the vector-search side of RAG.

Read paper → Read next: Dense Passage Retrieval for Open-Domain Question Answering →

Builder Bonus

ColBERT: Efficient and Effective Passage Search via Contextualized Late Interaction over BERT

Omar Khattab et al.

2020Must readIntermediate

Retrieval quality can improve a lot when you score token-level interactions instead of collapsing everything into one vector too early.

Why it matters

ColBERT is a strong builder paper because it explains why reranking and late interaction matter when plain dense retrieval hits quality limits.

What changed after this

Late interaction retrieval, stronger reranking, and hybrid retrieval stacks became more practical production choices.

Who should read

Retrieval engineers who need better search quality than a basic embedding index can provide.

Read paper → Read next: HyDE: Precise Zero-Shot Dense Retrieval without Relevance Labels →

Builder Bonus

RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval-Augmented Generation

Xisen Jin et al.

2024Must readIntermediate

Long-document RAG gets better when you build summaries and hierarchy into the retrieval layer itself.

Why it matters

RAPTOR made hierarchical retrieval concrete for builders dealing with long reports, books, and large document collections.

What changed after this

Tree-structured retrieval, recursive summarization, and long-document indexing became easier to reason about as product patterns.

Who should read

Builders working on long-form knowledge bases rather than short isolated passages.

Read paper → Read next: From Local to Global: A Graph RAG Approach to Query-Focused Summarization →

Builder Bonus

RAGAs: Reliable and General Evaluation for Retrieval-Augmented Generation

Xisen Jin et al.

2023Must readBeginner

A RAG system is only as trustworthy as your ability to evaluate retrieval and answer quality separately.

Why it matters

RAGAs helped make evaluation feel like a first-class part of the RAG stack instead of an afterthought handled by ad hoc spot checks.

What changed after this

RAG-specific metrics, retrieval-grounding checks, and automated eval loops became much more common in production teams.

Who should read

Anyone shipping RAG to users and needing something better than vibes to measure quality.

Read paper → Read next: Corrective Retrieval Augmented Generation →

Builder Bonus

RankGPT: An LLM As A Judge of Search Result Relevance

Zhiwei Han et al.

2023Should knowBeginner

LLMs are not only generators in a RAG stack; they can also help judge and rerank retrieval outputs.

Why it matters

RankGPT captures a very practical builder move: use the model to improve retrieval ordering before final generation.

What changed after this

LLM-based reranking, judge models, and more retrieval-aware prompt pipelines became easier to justify in production.

Who should read

Builders who care about squeezing more quality from the same index before touching the generator.

Read paper → Read next: Lost in the Middle: How Language Models Use Long Contexts →

How to read this page

A beginner, retrieval engineer, and product builder should not read the same subset in the same order. Pick the path that matches the kind of system you want to build.

Absolute beginner

Read only the shortest set that explains dense retrieval, the canonical RAG architecture, retrieval quality, adaptive RAG, and structured retrieval.

Dense Passage Retrieval for Open-Domain Question Answering Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks Lost in the Middle: How Language Models Use Long Contexts Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection From Local to Global: A Graph RAG Approach to Query-Focused Summarization

Builder path

Read the papers that explain how retrieval quality, context placement, repair loops, evaluation, and structured retrieval show up in real systems.

Dense Passage Retrieval for Open-Domain Question Answering FiD: Leveraging Passage Retrieval with Generative Models for Open-Domain Question Answering HyDE: Precise Zero-Shot Dense Retrieval without Relevance Labels REPLUG: Retrieval-Augmented Black-Box Language Models Lost in the Middle: How Language Models Use Long Contexts Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection Corrective Retrieval Augmented Generation RAGAs: Reliable and General Evaluation for Retrieval-Augmented Generation From Local to Global: A Graph RAG Approach to Query-Focused Summarization

Researcher path

Read the papers in order to see how dense retrieval, retrieval-aware language models, and modern adaptive RAG compound into the current field.

Dense Passage Retrieval for Open-Domain Question Answering REALM: Retrieval-Augmented Language Model Pre-Training Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks FiD: Leveraging Passage Retrieval with Generative Models for Open-Domain Question Answering RETRO: Improving language models by retrieving from trillions of tokens Atlas: Few-shot Learning with Retrieval Augmented Language Models HyDE: Precise Zero-Shot Dense Retrieval without Relevance Labels REPLUG: Retrieval-Augmented Black-Box Language Models Lost in the Middle: How Language Models Use Long Contexts Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection Corrective Retrieval Augmented Generation From Local to Global: A Graph RAG Approach to Query-Focused Summarization

Where to go after RAG

Once you understand RAG, the next question is usually what it serves: stronger base models, agent systems that use retrieval as memory, or richer multimodal knowledge interfaces.

LLM & Transformers

Read now

The model foundations behind the generators that RAG is trying to ground and extend.

AI Agents

Read now

How retrieval turns into memory, tool context, and environment grounding inside agent systems.

Multimodal AI

Coming next

How retrieval and grounding evolve once the knowledge base is not just text.

FAQ

Answers for the questions readers are most likely to search before or after landing on this page.

What is the most important RAG paper for beginners?

If you only read one paper, start with Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. It is the canonical paper that gave RAG its name and basic framing.

Is long context replacing RAG?

Not really. Lost in the Middle is a good reminder that longer context windows do not solve retrieval quality, ordering, or grounding on their own.

Why does DPR still matter if I use modern embedding APIs?

Because DPR explains the retrieval shift that made semantic passage search central. Even if you use a hosted embedding model today, the product logic still rests on the same retrieval idea.

What should I read after this page?

If you want stronger model foundations, go to the LLM guide. If you care about agent memory and tool grounding, go to AI Agents. If you care about richer knowledge interfaces, keep an eye on multimodal retrieval.

Want the daily version of this judgment?

This guide explains the long arc. Our daily feed explains what matters now.

Browse today’s papers → Browse all guides →