RAG Research: The Papers That Matter

RAG research that still shapes production systems is indexed here by the problem each paper addresses, not by citation count. Every entry names the published title, authors, year and arXiv identifier in prose, then routes to the page on this site that explains the mechanism. As of September 2026, titles and years are taken from the published record only; citation counts are never invented or copied from secondary lists. Gao et al.’s survey “Retrieval-Augmented Generation for Large Language Models: A Survey” (arXiv:2312.10997) still supplies the Naive / Advanced / Modular map most practitioners use to place a new paper. Prompting Guide’s RAG research overview and AWS’s “What is RAG” explainer both open on Lewis et al. 2020 as the naming paper, then fan into paradigms and systems. This index finishes the hop those lists stop at: after the citation, open the explainer that states what each stage does.

Foundational papers are the retrieve-then-generate baseline

Foundational RAG papers are the pre-2023 line that named the retrieve-then-generate recipe and the dense retriever underneath it. Read these before architecture variants, because later systems measure themselves against this baseline. Lewis et al. 2020 still appears as the naming paper on AWS’s RAG overview, Prompting Guide’s research page and Gao et al.’s survey, which is why this group opens the index rather than sitting inside a chronology.

The five entries below fix the parametric plus non-parametric split, retrieval inside pre-training, dual-encoder dense passage retrieval, Fusion-in-Decoder reading, and few-shot RAG at scale. Mechanism depth for dense retrieval lives under retrieval; the baseline pipeline shape lives under naive RAG. Years and identifiers are from the published arXiv records.

Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (Lewis et al., 2020; NeurIPS 2020; arXiv:2005.11401) introduces RAG as a seq2seq generator with a dense Wikipedia index as non-parametric memory, and shows gains on open-domain QA over parametric-only baselines. It is the naming paper for the method most production stacks still implement as retrieve, augment and generate.

REALM: Retrieval-Augmented Language Model Pre-Training (Guu et al., 2020; arXiv:2002.08909) puts retrieval inside pre-training so the model learns to fetch and use documents while it trains, not only at fine-tuning time. It is the precedent for treating retrieval as part of the model rather than a bolted-on prompt step, and it pairs with the baseline shape on naive RAG.

Dense Passage Retrieval for Open-Domain Question Answering (Karpukhin et al., 2020; arXiv:2004.04906) trains a dual-encoder dense retriever that embeds questions and passages into a shared space and retrieves by vector similarity. It remains the dense half of most modern RAG stacks; scoring and failure modes live under retrieval.

Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering (Izacard and Grave, 2020; arXiv:2007.01282) introduces Fusion-in-Decoder: encode retrieved passages independently, then fuse them in the decoder. The multi-passage reader pattern still shows up when a RAG pipeline packs several chunks into one generation step.

Atlas: Few-shot Learning with Retrieval Augmented Language Models (Izacard et al., 2022; arXiv:2208.03299) scales few-shot RAG by updating retriever and generator together so the system adapts with few examples. It bridges the 2020 recipe toward advanced RAG. Once the foundation is fixed, the next need is a map of the field rather than another primary architecture.

Survey papers map Naive, Advanced and Modular RAG

Survey papers are the field maps: they taxonomise paradigms and tell you which branch a new paper sits on. Use them to orient; use the architecture and evaluation groups below for the primary sources those surveys cite. Gao et al. 2023 remains the survey Prompting Guide and many GitHub paper tables still open with, because the Naive / Advanced / Modular cut matches how practitioners talk about control loops.

The six surveys below cover the general LLM-RAG map, broader AIGC framing, graph RAG, agentic RAG, evaluation methods and a 2025 systematic review of techniques and metrics. They do not replace reading a primary architecture paper when you are about to implement one pattern.

Retrieval-Augmented Generation for Large Language Models: A Survey (Gao et al., 2023; revised 2024; arXiv:2312.10997) publishes the widely used Naive / Advanced / Modular taxonomy, covering retrieval, generation and augmentation components plus evaluation. It is the survey to open first; paradigms live under RAG architectures.

Retrieval-Augmented Generation for AI-Generated Content: A Survey (Zhao et al., 2024; arXiv:2402.19473) widens the frame to how retrieval augments generation across modalities and tasks, with enhancement methods and benchmarks. Use it beside Gao when the question is what exists across AIGC, not how to wire one stack.

Graph Retrieval-Augmented Generation: A Survey (Peng et al., 2024; arXiv:2408.08921) indexes graph construction, graph-guided retrieval and graph-enhanced generation. Read it before implementing Graph RAG.

Agentic Retrieval-Augmented Generation: A Survey on Agentic RAG (Singh et al., 2025; arXiv:2501.09136) surveys agents that decide whether, what and how often to retrieve, including reflection, planning, tools and multi-agent setups. Map it to agentic RAG.

Evaluation of Retrieval-Augmented Generation: A Survey (Yu et al., 2024; arXiv:2405.07437) states what evaluating RAG uniquely requires versus plain generation or plain information retrieval. Pair it with the evaluation hub.

A Systematic Literature Review of Retrieval-Augmented Generation: Techniques, Metrics, and Challenges (Brown et al., 2025; arXiv:2508.06401) reviews techniques, metrics and open challenges across the 2020 to 2025 literature with explicit inclusion criteria. Use it when you need coverage of metrics rather than a single architecture bet. Surveys leave the question of which named control loop replaced plain retrieve-then-generate.

Architecture papers are control-loop patterns beyond naive RAG

Architecture papers name control-loop patterns that replace plain retrieve-then-generate. Each entry below is a primary source for a leaf under architectures: read the paper for the claim, the leaf for engineering trade-offs. Self-RAG, CRAG, RAPTOR, Microsoft GraphRAG, LightRAG, Modular RAG and Speculative RAG are the patterns ranking paper tables and Prompting Guide still treat as the post-naive line.

Google Research’s 2024 “sufficient context” line of work sits beside these patterns as a measurement of when more retrieved context helps or hurts; it is not a named architecture leaf on this map, but it explains why selective generation and quality gates appear inside Corrective and Self-RAG designs. Implementation depth stays on the architecture leaves, not in this index.

Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection (Asai et al., 2023; arXiv:2310.11511) trains a model with reflection tokens to decide when to retrieve and to critique its own passages and outputs. Retrieval is on demand, not on every query: Self-RAG.

Corrective Retrieval Augmented Generation (Yan et al., 2024; arXiv:2401.15884) grades retrieved documents and triggers corrective action, including web fallback, when retrieval looks wrong. It is retrieval with a quality gate: Corrective RAG.

RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval (Sarthi et al., 2024; arXiv:2401.18059) clusters and summarises chunks into a tree so retrieval can return a high-level summary or a leaf detail by question breadth. Hierarchical retrieval: RAPTOR.

From Local to Global: A Graph RAG Approach to Query-Focused Summarization (Edge et al., 2024; arXiv:2404.16130) is Microsoft GraphRAG: an entity graph plus community summaries for local and global questions over a corpus. Primary source for the graph-RAG leaf already linked from Peng et al. above.

LightRAG: Simple and Fast Retrieval-Augmented Generation (Guo et al., 2024; arXiv:2410.05779) uses dual-level graph retrieval aimed at lower indexing and query cost than heavy GraphRAG pipelines. Profiled at LightRAG.

Modular RAG: Transforming RAG Systems into LEGO-like Reconfigurable Frameworks (Gao et al., 2024; arXiv:2407.21059) treats the pipeline as swappable operators and topologies (linear, conditional, branching, looping) rather than one fixed chain. Framing page: Modular RAG.

Speculative RAG: Enhancing Retrieval Augmented Generation through Drafting (Wang et al., 2024; arXiv:2407.08223) has a smaller specialist model draft from subsets of retrieved documents while a larger generalist verifies and selects. Latency-aware drafting under the advanced-RAG leaf linked with Atlas above. Named architectures still leave the question of what changes the fetch step itself before any control loop runs.

Retrieval papers change what gets fetched before generation

Retrieval papers change what gets fetched before generation runs. They sit under retrieval and reranking; dense-passage foundations are listed above under foundational papers and are not duplicated as full entries here. HyDE and ColBERT are the two post-DPR levers ranking guides and GitHub tables still treat as must-read for query-time matching.

Hybrid dense-plus-lexical retrieval is the production default when identifiers and rare tokens matter; engineering treatment lives at hybrid search, with BM25 as the lexical baseline. Chunking research that changes what units enter the index, including late chunking, is owned by the chunking cluster rather than duplicated as a paper dump here.

Precise Zero-Shot Dense Retrieval without Relevance Labels (Gao et al., 2022; arXiv:2212.10496) introduces HyDE: generate a hypothetical answer, embed that, and retrieve against real documents, because the answer often sits closer to the corpus than the raw question. Mechanism: HyDE.

ColBERT: Efficient and Effective Passage Search via Contextualized Late Interaction over BERT (Khattab and Zaharia, 2020; arXiv:2004.12832) scores each query token by its best-matching document token (MaxSim) instead of one pooled embedding. Precision sits between bi-encoders and full cross-encoders: ColBERT.

Hybrid dense-plus-lexical retrieval fuses keyword and vector signals so rare tokens and paraphrases can both rank. The engineering treatment is at hybrid search, with BM25 as the lexical baseline at BM25. Vector databases such as Weaviate, Qdrant and Pinecone productise that hybrid path in production; this index only names the papers that changed the fetch. After fetch quality is named, the field still needs meters for whether the right chunk arrived and whether the answer stayed faithful.

Evaluation papers define how RAG quality is measured

Evaluation papers define how you measure whether the right chunk arrived and whether the answer stayed faithful to it. Start here for primary sources, then use the evaluation hub for metric formulas, tool comparisons and online monitoring. Paperguide’s RAG paper lists and Hugging Face’s RAG paper feed both keep Ragas and BEIR near the top of evaluation-related entries, which matches practitioner demand for automated faithfulness scoring and heterogeneous retrieval benchmarks.

Yu et al. 2024 under Survey papers remains the dedicated survey of RAG evaluation methods; it is cross-referenced rather than duplicated as a second full entry. Scores published only on vendor blogs are not restated here.

Ragas: Automated Evaluation of Retrieval Augmented Generation (Es et al., 2023; arXiv:2309.15217) scores faithfulness, context precision, context recall and answer relevance, largely with an LLM-as-judge. Profiled at Ragas; tool comparison at evaluation tools.

BEIR: A Heterogenous Benchmark for Zero-shot Evaluation of Information Retrieval Models (Thakur et al., 2021; arXiv:2104.08663) is a heterogeneous zero-shot IR benchmark still used to stress-test retrievers that feed RAG. Catalogue of public RAG benchmarks owns the practitioner list of datasets and leaderboards.

Google Research’s sufficient-context autorater work (as of 2024 to 2025 communications) measures when retrieved context is enough for a correct answer and when adding more context increases hallucinations, which is why selective generation appears beside faithfulness metrics in modern eval stacks. Depth on how to run those meters in production stays on the evaluation cluster. Meters without failure names still leave teams arguing about averages when the right evidence was retrieved and the answer is wrong.

Failure and robustness papers name production failure modes

Failure and robustness papers name a production failure mode rather than a new architecture. They matter because a system can retrieve the right evidence and still answer wrong. Flat paper lists on GitHub and Paperguide usually omit this divider; the failure taxonomy on this site is organised that way, so the index carries at least the position-bias primary source every utilization discussion still cites.

Hallucination and trustworthiness surveys exist in the wider literature; mechanism depth for those symptoms lives under failures rather than as an unverified citation ranking here. The entry below is the one position effect that still changes prompt packing after retrieval looks healthy.

Lost in the Middle: How Language Models Use Long Contexts (Liu et al., 2023; arXiv:2307.03172) measures how models under-use information placed in the middle of long contexts, even when the right passage was retrieved. Symptom page: lost in the middle; mitigation via position at context ordering. Broader symptom-to-stage routing lives under failure modes.

Comparison papers that ask RAG versus fine-tuning or long context are owned by the decisions cluster, not this index. Domain-specific application papers route to use cases. Embeddings and input-processing depth, past the late-chunking pointer above, route to embeddings and chunking. Framework and tool roundups route to decisions and evaluation tools. Past the citation, the hubs that implement these papers are architectures, retrieval, evaluation, the failure taxonomy, the pipeline map and the glossary.

RAG research that still shapes builds is a problem-grouped index: foundations that named the recipe, surveys that map the paradigms, architectures that changed the control loop, retrieval papers that changed the fetch, evaluation papers that named the meters, and failure papers that named utilization breaks after retrieval looked fine. Lewis et al. 2020 (arXiv:2005.11401) and Gao et al. 2023 (arXiv:2312.10997) remain the usual entry points. As of September 2026 the index cites titles, authors, years and arXiv identifiers only, never invented citation counts, and finishes each entry by routing to the explainer that owns the mechanism.