Agentic Retrieval-Augmented Generation: A Survey on Agentic RAG
by Singh, Ehtesham, Kumar, Talaei Khoei & Vasilakos (arXiv)
The taxonomy people are pointing at when they say "agentic RAG".
Overview
This survey by Aditi Singh (Cleveland State University), Abul Ehtesham (The Davey Tree Expert Company), Saket Kumar (The MathWorks), Tala Talaei Khoei (Northeastern University) and Athanasios V. Vasilakos was first posted to arXiv in January 2025 and last revised as v4 on 1 April 2026, and it is the standard map of the space. It builds in layers. First it traces RAG's evolution through the Naive, Advanced, Modular, Graph and Agentic paradigms, with a comparison table of key features, strengths and limitations for each. Then it covers agentic foundations — reflection, planning, tool use and multi-agent collaboration — and five workflow patterns: prompt chaining, routing, parallelisation, orchestrator-workers and evaluator-optimizer. The core contribution is a taxonomy of seven Agentic RAG architectures: the single-agent router, multi-agent systems with specialised retrievers, hierarchical agents, corrective RAG (a relevance grader that rewrites the query and retries), adaptive RAG (a classifier that skips retrieval, does one hop, or reasons over many), graph-based approaches via the Agent-G and GeAR frameworks, and agentic document workflows. A second table contrasts traditional RAG, Agentic RAG and document workflows across context maintenance, dynamic adaptability, orchestration, tool integration, scalability and reasoning complexity. It then works through applications in customer support, healthcare, legal, finance, education and multimodal graph workflows, names the implementation frameworks — LangChain, LlamaIndex, CrewAI, AutoGen, Amazon Bedrock, LlamaParse, Neo4j — and lists benchmarks including BEIR, MS MARCO, TREC, HotpotQA, MuSiQue and RAGBench before closing on open problems in evaluation, coordination overhead, memory management, efficiency and governance.
At a Glance
- Topic
- RAG
- Level
- Advanced
- Format
- Paper
- Cost
- Free
- Duration
- ~1.5-2 hours to read; survey with roughly 23 figures and 100+ references
- Provider
- Singh, Ehtesham, Kumar, Talaei Khoei & Vasilakos (arXiv)
- Hands-on
- No
- Certificate
- None
What You’ll Learn
- ✓Where each RAG paradigm breaks down, from naive through modular to graph
- ✓The seven Agentic RAG architectures and which query patterns each suits
- ✓How corrective RAG grades relevance and rewrites a failing query
- ✓How adaptive RAG classifies a query to skip or escalate retrieval
- ✓How graph-based retrieval enables multi-hop reasoning over knowledge bases
- ✓Which frameworks and benchmarks the field uses to compare RAG systems
- ✓The open problems: coordination overhead, memory, evaluation and governance
Highlights
- •The reference taxonomy people cite when they say "agentic RAG"
- •Actively maintained — revised to v4 in April 2026, not a stale 2025 preprint
- •Two comparison tables that map an architecture choice onto a workload
- •Roughly 23 figures and over 100 references, so it doubles as a reading list
- •Names concrete frameworks and benchmarks rather than staying abstract
Who It’s For
Best For
- ✓Engineers choosing an architecture before building a RAG system
- ✓Teams whose single-shot RAG pipeline fails on multi-hop questions
- ✓Researchers and technical leads needing a citable map of the field
Prerequisites
- •Working understanding of embeddings, vector retrieval and standard RAG
- •Comfort reading a densely cited academic survey
FAQ
What is Agentic Retrieval-Augmented Generation: A Survey on Agentic RAG?
The standard survey of agentic retrieval-augmented generation, revised to v4 in April 2026. It traces RAG from naive through modular and graph variants, defines seven agentic RAG architectures with comparison tables, works through six application domains, and names the frameworks and benchmarks the field uses. For engineers choosing an architecture before building, and for teams whose single-shot RAG fails on multi-hop questions.
Is Agentic Retrieval-Augmented Generation: A Survey on Agentic RAG free?
Agentic Retrieval-Augmented Generation: A Survey on Agentic RAG is free to access.
What level is Agentic Retrieval-Augmented Generation: A Survey on Agentic RAG for?
Agentic Retrieval-Augmented Generation: A Survey on Agentic RAG is aimed at a advanced audience. Recommended background: Working understanding of embeddings, vector retrieval and standard RAG, Comfort reading a densely cited academic survey.
How long does Agentic Retrieval-Augmented Generation: A Survey on Agentic RAG take?
Expect roughly ~1.5-2 hours to read; survey with roughly 23 figures and 100+ references. Most learners work through it at their own pace.
What will I learn from Agentic Retrieval-Augmented Generation: A Survey on Agentic RAG?
You'll learn: Where each RAG paradigm breaks down, from naive through modular to graph; The seven Agentic RAG architectures and which query patterns each suits; How corrective RAG grades relevance and rewrites a failing query; How adaptive RAG classifies a query to skip or escalate retrieval; How graph-based retrieval enables multi-hop reasoning over knowledge bases; Which frameworks and benchmarks the field uses to compare RAG systems; The open problems: coordination overhead, memory, evaluation and governance.
Topics
Sources
This page was written from 2 sources, 1 on domains other than arxiv.org.