Overview："Agentic Retrieval-Augmented Generation: A Comprehensive Survey"

Disclaimer: this is a report generated with my tool: https://github.com/DTeam-Top/tsw-cli. See it as an experiment not a formal research, 😄。

---

Mindmap

Summary

This survey paper explores Agentic Retrieval-Augmented Generation (Agentic RAG), a paradigm that enhances Large Language Models (LLMs) by integrating autonomous AI agents into the Retrieval-Augmented Generation (RAG) pipeline. It provides a detailed overview of the evolution of RAG, foundational principles of agentic intelligence, key applications across various industries, implementation strategies, and ethical considerations. The paper also identifies challenges in scaling Agentic RAG systems and suggests future research directions.

Terminology

• Large Language Models (LLMs): AI models trained on vast amounts of text data, capable of generating human-like text and understanding natural language.
• Retrieval-Augmented Generation (RAG): A framework that combines the generative capabilities of LLMs with external retrieval mechanisms to provide contextually relevant and up-to-date responses.
• Agentic RAG: An advanced RAG paradigm that incorporates autonomous AI agents into the RAG pipeline to dynamically manage retrieval strategies, refine contextual understanding, and adapt workflows.
• Agentic Patterns: Design methodologies, including reflection, planning, tool use, and multi-agent collaboration, that guide the behavior of agents in Agentic RAG systems.

Main Points

Evolution of RAG Paradigms

The paper traces the evolution of RAG from basic keyword-based retrieval to sophisticated agent-based systems.

• Naïve RAG: The foundational implementation using keyword-based retrieval techniques like TF-IDF and BM25. Limited by contextual awareness and scalability issues.
• Advanced RAG: Incorporates semantic understanding and dense retrieval models like DPR to improve precision and nuanced understanding.
• Modular RAG: Emphasizes flexibility and customization through independent, reusable components and hybrid retrieval strategies.
• Graph RAG: Integrates graph-based data structures to enhance multi-hop reasoning and contextual enrichment.
• Agentic RAG: Introduces autonomous agents for dynamic decision-making, iterative refinement, and adaptive retrieval strategies.

Core Principles of Agentic Intelligence

Agentic intelligence is the foundation of Agentic RAG, enabling dynamic decision-making and adaptability.

• Components of an AI Agent: LLM with defined roles, short-term and long-term memory, planning capabilities, and access to external tools.
• Agentic Patterns: Reflection (self-evaluation), planning (task decomposition), tool use (external resources), and multi-agent collaboration (specialized agents working together).

Agentic Workflow Patterns

Agentic workflow patterns structure LLM-based applications to optimize performance and accuracy.

• Prompt Chaining: Decomposes tasks into sequential steps.
• Routing: Directs inputs to specialized processes based on classification.
• Parallelization: Executes independent processes concurrently.
• Orchestrator-Workers: Uses a central orchestrator to delegate subtasks to worker models.
• Evaluator-Optimizer: Iteratively refines output based on feedback.

Taxonomy of Agentic RAG Systems

The paper categorizes Agentic RAG systems into architectural frameworks.

• Single-Agent Agentic RAG: A centralized system where one agent manages retrieval, routing, and integration of information.
• Multi-Agent Agentic RAG: Distributes responsibilities across multiple specialized agents for complex workflows.
• Hierarchical Agentic RAG: Employs a multi-tiered structure with higher-level agents overseeing lower-level agents for strategic decision-making.
• Agentic Corrective RAG: Implements mechanisms for self-correction of retrieval results.
• Adaptive Agentic RAG: Dynamically adjusts query handling strategies based on query complexity.
• Graph-Based Agentic RAG:
  • Agent-G: Combines graph knowledge bases with unstructured document retrieval.
  • GeAR: Enhances traditional RAG by incorporating graph expansion techniques.
• Agentic Document Workflows (ADW): Automates complex document-centric processes, integrating document parsing, retrieval, reasoning, and structured outputs with intelligent agents.

Improvements And Creativity

The survey provides a structured and detailed overview of Agentic RAG, offering a clear taxonomy of architectures, applications, and tools. The comparative analysis of different RAG paradigms and frameworks is particularly insightful. The paper also addresses the ethical considerations and challenges in scaling these systems, providing a balanced perspective.

Insights

Agentic RAG represents a significant advancement in AI, offering enhanced adaptability and precision compared to traditional RAG systems. The integration of autonomous agents enables more complex, real-time, and context-aware applications. Future research should focus on addressing the challenges of coordination complexity, scalability, and ethical considerations to fully realize the potential of Agentic RAG. The development of specialized benchmarks and datasets is also crucial for evaluating agentic capabilities.

References

• Source 8 - Agent-g: An agentic framework for graph retrieval augmented generation
• Source 12 - Building effective agents
• Source 13 - Langgraph workflows tutorial
• Source 14 - Agentic retrieval-augmented generation for time series analysis
• Source 16 - Graph retrieval-augmented generation: A survey
• Source 17 - Revolutionizing mental health care through langchain: A journey with a large language model
• Source 20 - Retrieval-augmented generation for large language models: A survey
• Source 30 - What is agentic rag?
• Source 31 - Corrective retrieval augmented generation
• Source 32 - Langgraph crag: Contextualized retrieval-augmented generation tutorial
• Source 33 - Adaptive-rag: Learning to adapt retrieval-augmented large language models through question complexity
• Source 34 - Langgraph adaptive rag: Adaptive retrieval-augmented generation tutorial
• Source 35 - Gear: Graph-enhanced agent for retrieval-augmented generation
• Source 36 - Introducing agentic document workflows
• Source 37 - How twitch used agentic workflow with rag on amazon bedrock to supercharge ad sales
• Source 38 - Patient case summary workflow using llamacloud
• Source 39 - Contract review workflow using llamacloud
• Source 40 - Auto insurance claims workflow using llamacloud
• Source 41 - Research paper report generation workflow using llamacloud
• Source 42 - Langgraph agentic rag: Nodes and edges tutorial
• Source 43 - Agentic rag with llamaindex
• Source 44 - Agentic rag: Turbocharge your retrieval-augmented generation with query reformulation and self-query
• Source 45 - Agentic rag: Combining rag with agents for enhanced information retrieval
• Source 46 - CrewAI: A github repository for ai projects
• Source 47 - AG2: A github repository for advanced generative ai research
• Source 48 - Autogen: Enabling next-gen llm applications via multi-agent conversation framework
• Source 50 - Swarm: Lightweight multi-agent orchestration framework
• Source 51 - Agentic rag using vertex ai
• Source 52 - Semantic kernel overview
• Source 53 - Semantic kernel github repository
• Source 54 - Agentic rag: Ai agents with ibm granite models
• Source 55 - Beir: A heterogenous benchmark for zero-shot evaluation of information retrieval models
• Source 56 - Ms marco: A human generated machine reading comprehension dataset
• Source 57 - Overview of the trec 2022 deep learning track
• Source 58 - Musique: Multihop questions via single-hop question composition
• Source 59 - Constructing a multi-hop qa dataset for comprehensive evaluation of reasoning steps
• Source 60 - Hotpotqa: A dataset for diverse, explainable multi-hop question answering
• Source 61 - Ragbench: Explainable benchmark for retrieval-augmented generation systems
• Source 62 - Bergen: A benchmarking library for retrieval-augmented generation
• Source 63 - Flashrag: A modular toolkit for efficient retrieval-augmented generation research
• Source 64 - Gnn-rag: Graph neural retrieval for large language model reasoning
• Source 65 - Natural questions: A benchmark for question answering research
• Source 66 - Triviaqa: A large scale distantly supervised challenge dataset for reading comprehension
• Source 67 - Squad: 100,000+ questions for machine comprehension of text
• Source 68 - Semantic parsing on freebase from question-answer pairs
• Source 69 - When not to trust language models: Investigating effectiveness of parametric and non-parametric memories
• Source 70 - Eli5: Long form question answering
• Source 71 - The narrativeqa reading comprehension challenge
• Source 72 - Asqa: Factoid questions meet long-form answers
• Source 73 - QMSum: A new benchmark for query-based multi-domain meeting summarization
• Source 74 - A dataset of information-seeking questions and answers anchored in research papers
• Source 75 - COVID-QA: A question answering dataset for COVID-19
• Source 76 - Cmb: A comprehensive medical benchmark in chinese
• Source 77 - Quality: Question answering with long input texts, yes!
• Source 78 - CommonsenseQA: A question answering challenge targeting commonsense knowledge
• Source 79 - G-retriever: Retrieval-augmented generation for textual graph understanding and question answering
• Source 80 - Document-level event argument extraction by conditional generation
• Source 81 - Multi-sentence argument linking
• Source 82 - Wizard of wikipedia: Knowledge-powered conversational agents
• Source 83 - Large language models as source planner for personalized knowledge-grounded dialogue
• Source 84 - Long time no see! open-domain conversation with long-term persona memory
• Source 85 - Conditional generation and snapshot learning in neural dialogue systems
• Source 86 - Ups and downs: Modeling the visual evolution of fashion trends with one-class collaborative filtering
• Source 87 - HellaSwag: Can a machine really finish your sentence?
• Source 88 - The cot collection: Improving zero-shot and few-shot learning of language models via chain-of-thought fine-tuning
• [Source 89(https://arxiv.org/abs/1806.01432) - Complex sequential question answering: Towards learning to converse over linked question answer pairs with a knowledge graph
• Source 90 - FEVER: a large-scale dataset for fact extraction and VERification
• Source 91 - Explainable automated fact-checking for public health claims
• Source 92 - Did aristotle use a laptop? a question answering benchmark with implicit reasoning strategies
• Source 93 - Wikiasp: A dataset for multi-domain aspect-based summarization
• Source 94 - Don’t give me the details, just the summary! topic-aware convolutional neural networks for extreme summarization
• Source 95 - Recursive deep models for semantic compositionality over a sentiment treebank
• Source 96 - Vio-lens: A novel dataset of annotated social network posts leading to different forms of communal violence and its evaluation
• Source 97 - Codesearchnet challenge: Evaluating the state of semantic code search
• Source 98 - "knowing when you don’t know": A multilingual relevance assessment dataset for robust retrieval-augmented generation
• Source 99 - Pointer sentinel mixture models
• Source 100 - Training verifiers to solve math word problems
• Source 101 - The JRC-Acquis: A multilingual aligned parallel corpus with 20+ languages

Paper: Agentic Retrieval-Augmented Generation: A Survey on Agentic RAG

---

Report generated by TSW-X
Advanced Research Systems Division
Date: 2025-03-05 09:06:41.582401