1. Bibliographic Information

1.1. Title

The title of the paper is "Agentic Retrieval-Augmented Generation: A Survey on Agentic RAG".

1.2. Authors

The authors are:

• Aditi Singh (Department of Computer Science, Cleveland State University, Cleveland, OH, USA)
• Abul Ehtesham (The Davey Tree Expert Company, Kent, OH, USA)
• Saket Kumar (The MathWorks Inc, Natick, MA, USA)
• Tala Talaei Khoei (Khoury College of Computer Science, Roux Institute at Northeastern University, Portland, ME, USA)

1.3. Journal/Conference

This paper is a survey, published as a preprint. The abstract states "Published at (UTC): 2025-01-15T20:40:25.000Z". While it doesn't explicitly state a journal or conference, survey papers are often published in dedicated survey journals or presented at major conferences. Given its publication on arXiv, it is likely awaiting or has been submitted for peer review. ArXiv is a widely recognized platform for preprints in various scientific fields, allowing early dissemination of research.

1.4. Publication Year

The publication timestamp indicates a release in 2025 (specifically, January 15, 2025).

1.5. Abstract

Large Language Models (LLMs) have transformed AI with their human-like text generation and understanding capabilities. However, their reliance on static training data leads to outdated or inaccurate responses for real-time queries. Retrieval-Augmented Generation (RAG) addresses this by integrating real-time data retrieval to provide contextually relevant and current information. Traditional RAG systems, however, are limited by static workflows and struggle with multi-step reasoning and complex task management.

Agentic Retrieval-Augmented Generation (Agentic RAG) overcomes these limitations by embedding autonomous AI agents within the RAG pipeline. These agents employ agentic design patterns such as reflection, planning, tool use, and multi-agent collaboration to dynamically manage retrieval strategies, iteratively refine contextual understanding, and adapt workflows to complex tasks. This integration enhances Agentic RAG systems with flexibility, scalability, and context awareness across various applications.

This survey provides a comprehensive overview of Agentic RAG, covering its foundational principles, the evolution of RAG, a detailed taxonomy of Agentic RAG architectures, key applications (e.g., healthcare, finance, education), practical implementation strategies, and challenges related to scaling, ethical decision-making, and performance optimization. It also offers insights into relevant frameworks and tools.

1.6. Original Source Link

The original source link is https://arxiv.org/abs/2501.09136. The PDF link is https://arxiv.org/pdf/2501.09136v3.pdf. This paper is published as a preprint on arXiv.

2. Executive Summary

2.1. Background & Motivation

The core problem the paper aims to solve stems from the inherent limitations of Large Language Models (LLMs). While LLMs excel at human-like text generation and natural language understanding, their knowledge is static, confined to the data they were trained on. This leads to outdated, inaccurate, or hallucinated outputs when confronted with dynamic, real-time queries or tasks requiring up-to-date information.

Retrieval-Augmented Generation (RAG) emerged as a solution by integrating external, real-time data sources into the LLM generation process. This significantly improved the factual accuracy and relevance of LLM responses. However, traditional RAG systems are often characterized by static workflows. They struggle with multi-step reasoning, lack adaptability to complex task requirements, and cannot dynamically manage retrieval strategies or iteratively refine their contextual understanding. This limitation hinders their effectiveness in real-world applications that demand dynamic decision-making and flexible task management.

The paper's entry point is the recognition that combining RAG with autonomous AI agents can address these shortcomings. The innovative idea is to embed AI agents – entities capable of perceiving, reasoning, planning, acting, and collaborating – into the RAG pipeline. This agentic approach is hypothesized to inject the necessary dynamic adaptability and intelligent orchestration required for RAG systems to handle highly complex and evolving tasks.

2.2. Main Contributions / Findings

The primary contributions of this survey paper are:

• Comprehensive Exploration of Agentic RAG: It provides a foundational understanding of Agentic RAG, detailing its principles and tracing its evolution from earlier RAG paradigms (Naïve, Advanced, Modular, Graph RAG).

• Detailed Taxonomy of Architectures: The paper presents a structured taxonomy of Agentic RAG architectures, including Single-Agent, Multi-Agent, Hierarchical, Corrective, Adaptive, and Graph-Based Agentic RAG systems, as well as Agentic Document Workflows (ADW). Each is described with its workflow, features, advantages, and challenges.

• Identification of Agentic Design Patterns: It highlights key agentic design patterns (reflection, planning, tool use, multi-agent collaboration) that enable Agentic RAG systems to manage dynamic workflows and complex problem-solving.

• Overview of Applications: The survey showcases key applications of Agentic RAG across diverse industries like healthcare, finance, education, legal, and customer support, demonstrating its transformative potential.

• Practical Implementation Strategies: It examines practical implementation strategies, discussing frameworks and tools (LangChain, LlamaIndex, CrewAI, AutoGen, Semantic Kernel, etc.) for building Agentic RAG systems.

• Discussion of Challenges and Future Directions: The paper addresses critical challenges in scaling, ethical decision-making, and performance optimization, while also outlining future research directions for the field.

  The key conclusion is that Agentic RAG represents a paradigm shift that significantly enhances the capabilities of LLMs by integrating autonomous agents into the RAG pipeline. These systems deliver unparalleled flexibility, scalability, and context-awareness, positioning them as a cornerstone for next-generation AI applications that can tackle complex, dynamic, and knowledge-intensive challenges that traditional RAG systems cannot. The findings emphasize that Agentic RAG moves beyond static workflows to provide dynamic and adaptive responses, ultimately overcoming the contextual integration, multi-step reasoning, and scalability limitations of previous RAG approaches.

3. Prerequisite Knowledge & Related Work

3.1. Foundational Concepts

To understand Agentic RAG, a novice reader should first grasp the following foundational concepts:

• Large Language Models (LLMs): LLMs are a type of artificial intelligence model trained on vast amounts of text data to understand, generate, and process human language. They can perform various tasks like answering questions, summarizing text, translating languages, and writing creative content. Examples include OpenAI's GPT-4, Google's PaLM, and Meta's LLaMA. Their core limitation, as highlighted in the paper, is their reliance on static training data, meaning their knowledge is fixed at the time of training and can become outdated.
• Natural Language Understanding (NLU): A subfield of AI that focuses on enabling computers to understand human language as it is spoken and written. NLU is crucial for LLMs to interpret user queries and context.
• Generative AI: A category of AI models that can generate new content, such as text, images, audio, or video, based on patterns learned from their training data. LLMs are a prominent form of generative AI.
• Retrieval-Augmented Generation (RAG): RAG is an AI framework that enhances the factual accuracy and up-to-dateness of LLMs by giving them access to external, up-to-date knowledge bases. Instead of solely relying on their pre-trained knowledge, RAG systems retrieve relevant information from a separate data source (e.g., a database, document corpus, or the internet) and then augment the LLM's prompt with this information before generating a response. This mitigates the LLM's tendency to hallucinate (generate factually incorrect but plausible-sounding information) and provides contextual relevance.
• AI Agents (Agentic Intelligence): In AI, an agent is an autonomous entity that perceives its environment through sensors and acts upon that environment through effectors (actions). Agentic intelligence refers to the ability of these AI agents to reason, plan, learn, and autonomously perform tasks to achieve specific goals, often interacting with their environment and other agents. Key components of an AI agent typically include an LLM as its reasoning engine, memory (short-term for current context, long-term for accumulated knowledge), planning capabilities (for breaking down tasks and strategizing), and tools (for interacting with external systems or data).
• Vector Databases: These are specialized databases designed to store and query vector embeddings (numerical representations of data like text or images) efficiently. They are fundamental to RAG systems for semantic search, allowing the system to find documents whose meaning is similar to a user's query, rather than just matching keywords.
• APIs (Application Programming Interfaces): APIs are sets of rules and protocols for building and interacting with software applications. In RAG and Agentic RAG, APIs allow agents or LLMs to interact with external services, databases, or specialized tools (e.g., weather data, financial market data, translation services).

3.2. Previous Works

The paper frames Agentic RAG as the latest evolution of RAG paradigms, building upon several prior approaches. It implicitly or explicitly references foundational elements and challenges addressed by these earlier works:

• Naive RAG: This is the simplest form of RAG, relying on basic keyword-based retrieval methods like TF-IDF (Term Frequency-Inverse Document Frequency) or BM25 (Best Match 25).

  • TF-IDF: A numerical statistic that reflects how important a word is to a document in a corpus. It increases proportionally to the number of times a word appears in the document but is offset by the frequency of the word in the corpus, which helps to adjust for the fact that some words appear more frequently in general.
  • BM25: A ranking function used by search engines to estimate the relevance of documents to a given search query. It is a bag-of-words model that ranks a set of documents based on the query terms appearing in each document, regardless of the inter-relationship between the query terms within a document.
  • Limitation: Naive RAG suffers from a lack of contextual awareness (relies on lexical matching, not semantic understanding), fragmented outputs, and scalability issues with large datasets due to its keyword-based nature.

• Advanced RAG: This paradigm improved upon Naive RAG by incorporating semantic understanding and contextual awareness.

  • Dense Retrieval Models: These models, such as Dense Passage Retrieval (DPR), represent queries and documents as high-dimensional vector embeddings. Similarity between query and documents is then measured by vector distance (e.g., cosine similarity).
  • DPR (Dense Passage Retrieval): A method that uses neural networks to embed queries and documents into dense vector representations. Retrieval is performed by finding documents whose vector embeddings are closest to the query's vector embedding.
  • Contextual Re-Ranking: Neural models are used to re-rank the initially retrieved documents, prioritizing those most contextually relevant to the query, even if their initial semantic similarity score wasn't the highest.
  • Iterative Retrieval (Multi-hop Retrieval): Introduces mechanisms to perform multiple retrieval steps, allowing the system to reason across several documents to answer complex multi-hop queries (queries requiring information from more than one source).
  • Limitation: Despite advancements, Advanced RAG still faced computational overhead and limited scalability for very large datasets or complex, multi-step queries.

• Modular RAG: This evolution focused on flexibility and customization by breaking down the RAG pipeline into independent, reusable components.

  • Hybrid Retrieval Strategies: Combines sparse retrieval (like BM25) with dense retrieval (like DPR) to leverage the strengths of both, improving accuracy across diverse query types.
  • Tool Integration: Incorporates external APIs, databases, or computational tools for specialized tasks, moving beyond just document retrieval.
  • Composable Pipelines: Allows retrievers, generators, and other components to be replaced or reconfigured independently.
  • Limitation: While offering great flexibility, Modular RAG still relied on predefined workflows and lacked the dynamic adaptability of true autonomous agents.

• Graph RAG: This paradigm integrated graph-based knowledge structures into RAG to enhance reasoning over relationships between entities.

  • Node Connectivity: Captures and reasons over relationships within structured data (e.g., a knowledge graph linking concepts, people, and events).
  • Hierarchical Knowledge Management: Manages both structured and unstructured data by leveraging graph hierarchies.
  • Context Enrichment: Adds relational understanding by traversing graph pathways.
  • Limitation: Graph RAG faced challenges with scalability (especially with extensive graph sources), data dependency (requiring high-quality graph data), and complexity of integration with unstructured retrieval systems.

• Agentic Design Patterns: The paper specifically mentions reflection, planning, tool use, and multi-agent collaboration as crucial for Agentic RAG. These patterns are often explored in agent-based systems literature (e.g., in works like Self-Refine [27], Reflexion [28], CRITIC [23] for reflection; various works on LLM planning [24]; function calling and tool integration in LLMs; and multi-agent systems [29] like AutoGen [48]). The paper integrates these as core elements that allow Agentic RAG to transcend the limitations of previous RAG paradigms.

3.3. Technological Evolution

The evolution of RAG from Naive RAG to Agentic RAG can be seen as a continuous effort to make LLM responses more accurate, contextually relevant, adaptive, and capable of complex reasoning in dynamic environments.

1. Phase 1: Basic Augmentation (Naive RAG): Early RAG focused on simply fetching relevant documents using keyword matching and feeding them to the LLM. This was a significant step to overcome static knowledge but lacked sophistication.
2. Phase 2: Semantic and Iterative Improvement (Advanced RAG): The introduction of dense vector search and re-ranking brought semantic understanding into play, making retrieval more intelligent. Multi-hop retrieval hinted at more complex reasoning capabilities.
3. Phase 3: Flexibility and External Interaction (Modular RAG): Recognizing the need for customization, Modular RAG allowed for hybrid approaches and tool integration, expanding the LLM's reach beyond just document corpora to other APIs and computational tools.
4. Phase 4: Relational Reasoning (Graph RAG): To address hallucinations and improve reasoning over structured data, Graph RAG introduced knowledge graphs, enabling the LLM to leverage explicit relationships.
5. Phase 5: Autonomous and Dynamic Orchestration (Agentic RAG): The current frontier, where autonomous AI agents are embedded. This is a leap from predefined, static pipelines to systems that can dynamically adapt, plan complex tasks, iteratively refine, and collaborate to achieve goals. It's about injecting intelligence and adaptability into the RAG pipeline itself, rather than just improving its retrieval or generation components in isolation.

3.4. Differentiation Analysis

Compared to the main methods in related work, Agentic RAG introduces several core differentiators:

• Dynamic Decision-Making vs. Static Workflows:

  • Traditional RAG (Naive, Advanced, Modular, Graph) largely relies on static, predefined workflows. The retrieval strategy, re-ranking steps, and integration points are typically fixed or configured manually.
  • Agentic RAG introduces autonomous agents that can dynamically evaluate queries, select optimal retrieval strategies, choose appropriate tools, and adapt workflows in real-time based on the task's demands and intermediate results. This is a fundamental shift from fixed pipelines to intelligent, self-organizing systems.

• Iterative Refinement and Self-Correction:

  • While Advanced RAG introduced re-ranking and multi-hop retrieval, the feedback loops were often implicit or limited.
  • Agentic RAG explicitly incorporates iterative refinement through reflection and self-critique patterns. Agents can evaluate their own outputs, identify shortcomings, and refine their approach, leading to higher accuracy and relevance over multiple steps.

• Complex Multi-Step Reasoning and Task Management:

  • Traditional RAG struggles with complex multi-step queries that require information synthesis across diverse sources or multiple reasoning steps.
  • Agentic RAG excels here by leveraging planning capabilities to break down complex problems into manageable sub-tasks. Multi-agent collaboration further enables specialized agents to work together on different aspects of a complex task, synthesizing their findings for a comprehensive response.

• Enhanced Tool Use and External Interaction:

  • Modular RAG brought tool integration, but the selection and orchestration of tools were still largely programmatic.
  • Agentic RAG empowers agents to autonomously select and utilize tools (like APIs, databases, or web search) as needed, making the system more versatile and capable of interacting with the real world beyond simple data retrieval.

• Scalability for Multi-Domain and Dynamic Tasks:

  • Traditional RAG often faces scalability issues when dealing with highly dynamic data or diverse knowledge domains.

  • Agentic RAG, especially with multi-agent and hierarchical architectures, is designed for scalability in complex, multi-domain applications by distributing tasks and allowing specialized agents to handle specific knowledge sources or processing types.

    In essence, the core innovation of Agentic RAG is the infusion of proactive intelligence and adaptability into the RAG pipeline, transforming it from a reactive data retrieval and generation mechanism into an autonomous, problem-solving system.

4. Methodology

The paper describes Agentic RAG as a paradigm shift that embeds autonomous AI agents into the RAG pipeline to overcome the limitations of traditional RAG systems. The core methodology revolves around leveraging agentic design patterns (reflection, planning, tool use, multi-agent collaboration) to dynamically manage retrieval strategies, iteratively refine contextual understanding, and adapt workflows. The paper outlines several agentic workflow patterns and then categorizes Agentic RAG systems into a detailed taxonomy.

4.1. Core Principles of Agentic Intelligence

The foundation of Agentic RAG lies in Agentic Intelligence, where AI agents are intelligent entities capable of perceiving, reasoning, and autonomously performing tasks.

4.1.1. Components of an AI Agent

An AI agent comprises the following key components, as illustrated in Figure 7:

• LLM (with defined Role and Task): This serves as the agent's primary reasoning engine and dialogue interface. It interprets user queries, generates responses, and maintains coherence. The LLM is guided by a specific role and task assigned to the agent.

• Memory (Short-Term and Long-Term): Memory captures context and relevant data across interactions.

  • Short-term memory tracks the immediate conversation state and current context.
  • Long-term memory stores accumulated knowledge and agent experiences, enabling learning and recall over time.

• Planning (Reflection & Self-Critique): This guides the agent's iterative reasoning process. Through reflection, query routing, or self-critique, the agent can break down complex tasks, monitor progress, and refine its actions.

• Tools (Vector Search, Web Search, APIs, etc.): Tools expand the agent's capabilities beyond text generation. They enable access to external resources, real-time data, or specialized computations (e.g., databases, APIs for external services, web search engines for up-to-date information).

  The following figure (Figure 7 from the original paper) shows an overview of AI agents:

  Figure 7: An Overview of AI Agents

4.1.2. Agentic Design Patterns

Agentic patterns provide structured methodologies that guide the behavior of agents, enabling them to dynamically adapt, plan, and collaborate within Agentic RAG systems.

4.1.2.1. Reflection

Reflection is a meta-cognitive process where agents evaluate their performance, outputs, or reasoning steps to identify areas for improvement. It enhances coherence and accuracy across tasks. Agents can iteratively refine their outputs by critically evaluating their retrieval results or generated text. In multi-agent systems, reflection can involve distinct roles, such as one agent generating an output while another critiques it, fostering collaborative improvement. Reflection is shown to significantly improve performance in studies like Self-Refine, Reflexion, and CRITIC.

The following figure (Figure 8 from the original paper) shows an overview of agentic self-reflection:

Figure 8: An Overview of Agentic Self- Reflection

4.1.2.2. Planning

Planning involves the agent's ability to devise a sequence of actions or strategies to achieve a goal. Agentic planning allows LLMs to respond to dynamic and uncertain scenarios by decomposing complex tasks into smaller, manageable sub-tasks. This adaptability enables agents to handle tasks that cannot be entirely predefined, bringing flexibility to decision-making. Planning helps structure tasks requiring adaptive, step-by-step workflows.

The following figure (Figure 9 from the original paper) shows an overview of agentic planning and tool use (left side):

Figure 9: Overview of Agentic Planning and Tool Use

4.1.2.3. Tool Use

Tool use extends an agent's capabilities beyond its inherent LLM knowledge by enabling it to interact with external resources. Agents can leverage APIs and various tools (e.g., web search, calculators, databases) to retrieve real-time data, perform specific computations, or access domain-specific information. This significantly enhances their operational workflow and allows them to provide more accurate and contextually relevant outputs. The ability to autonomously select and execute tools is a critical aspect of advanced agentic workflows.

The following figure (Figure 9 from the original paper) shows an overview of agentic planning and tool use (right side):

Figure 9: Overview of Agentic Planning and Tool Use

4.1.2.4. Multi-Agent

Multi-agent systems involve multiple AI agents that communicate and share intermediate results to achieve a common goal. This approach enhances the overall workflow by distributing tasks, enabling specialization, and improving adaptability to complex problems. Multi-agent systems allow for decomposing intricate tasks into smaller, manageable sub-tasks, assigned to different agents. Each agent operates with its own memory and workflow, contributing to a collaborative problem-solving process. Frameworks like AutoGen, Crew AI, and LangGraph facilitate the implementation of effective multi-agent solutions.

The following figure (Figure 10 from the original paper) shows an overview of multi-agent systems:

Figure 10: An Overview of MultiAgent

4.2. Agentic Workflow Patterns: Adaptive Strategies for Dynamic Collaboration

These patterns define how agents interact and structure their tasks, enabling LLMs to handle complex queries efficiently.

4.2.1. Prompt Chaining: Enhancing Accuracy Through Sequential Processing

Prompt chaining decomposes a complex task into multiple sequential steps, where each step builds upon the previous one. This enhances accuracy by allowing for step-by-step reasoning but can increase latency.

The following figure (Figure 11 from the original paper) illustrates the prompt chaining workflow:

Figure 11: Illustration of Prompt Chaining Workflow

When to Use: This pattern is suitable for tasks requiring an ordered sequence of operations, where the output of one step serves as input for the next, ensuring accuracy. Example Applications:

• Generating marketing content in one language, then translating it into another while preserving nuances.
• Structuring document creation by first generating an outline, verifying its completeness, and then developing the full text.

4.2.2. Routing: Directing Inputs to Specialized Processes

Routing automatically directs different types of inputs to specialized agents or processes based on their characteristics. This ensures that distinct queries or tasks are handled by the most appropriate components, improving efficiency and response quality.

The following figure (Figure 12 from the original paper) illustrates the routing workflow:

Figure 12: Illustration Routing Workflow

When to Use: Ideal when handling diverse types of queries that require different processing pathways or tools, optimizing performance for each category. Example Applications:

• Directing customer service queries into categories like technical support, refund requests, or general inquiries.
• Assigning simple queries to smaller models for cost efficiency, while complex requests go to advanced models.

4.2.3. Parallelization: Speeding Up Processing Through Concurrent Execution

Parallelization divides a task into independent processes that run simultaneously. This reduces latency and can enhance reliability by cross-checking results from multiple processes.

The following figure (Figure 13 from the original paper) illustrates the parallelization workflow:

Figure 13: Illustration of Parallelization Workflow

When to Use: Useful when tasks can be executed independently to enhance speed or when multiple outputs can be used to improve confidence (e.g., through voting). Example Applications:

• Sectioning: Splitting tasks like content moderation, where one model screens input while another generates a response.
• Voting: Using multiple models to cross-check code for vulnerabilities or analyze content moderation decisions.

4.2.4. Orchestrator-Workers: Dynamic Task Delegation

This workflow involves an orchestrator agent that dynamically assigns tasks to worker agents. The orchestrator manages the overall flow, breaks down complex problems, and delegates sub-tasks, while worker agents execute specific functions. This is particularly useful for tasks that require dynamic decomposition and real-time adaptation.

The following figure (Figure 14 from the original paper) illustrates the orchestrator-workers workflow:

Figure 14: Ilustration of Orchestrator-Workers Workflow

When to Use: Best suited for tasks requiring dynamic decomposition and real-time adaptation, especially when sub-tasks are not predefined. Example Applications:

• Automatically modifying multiple files in a codebase based on the nature of requested changes.
• Conducting real-time research by gathering and synthesizing relevant information from multiple sources.

4.2.5. Evaluator-Optimizer: Refining Output Through Iteration

The evaluator-optimizer pattern involves a generator agent creating an output, which is then assessed by an evaluator agent. Based on feedback from the evaluator, the generator refines its output iteratively until a satisfactory result is achieved.

The following figure (Figure 15 from the original paper) illustrates the evaluator-optimizer workflow:

Figure 15: Illustration of Evaluator-Optimizer Workflow

When to Use: Effective when iterative refinement significantly enhances response quality, especially when clear evaluation criteria exist. Example Applications:

• Improving literary translations through multiple evaluation and refinement cycles.
• Conducting multi-round research queries where additional iterations refine search results.

4.3. Taxonomy of Agentic RAG Systems

Agentic Retrieval-Augmented Generation (RAG) systems can be categorized into distinct architectural frameworks, each with unique strengths and limitations.

4.3.1. Single-Agent Agentic RAG: Router

A Single-Agent Agentic RAG system, often acting as a router, serves as a centralized decision-making system where a single agent manages the entire retrieval and generation process. This agent is responsible for evaluating the query, selecting appropriate tools or data sources, and synthesizing the final response.

The following figure (Figure 16 from the original paper) shows an overview of single agentic RAG:

Figure 16: An Overview of Single Agentic RAG

Workflow

1. Query Submission and Evaluation: A user submits a query, which is received by a coordinating agent (or master retrieval agent). This agent analyzes the query to determine the most suitable sources of information.
2. Knowledge Source Selection: Based on the query's type, the coordinating agent chooses from various retrieval options:
   • Structured Databases: For tabular data, it may use a Text-to-SQL engine (e.g., for PostgreSQL or MySQL).
   • Semantic Search: For unstructured information (e.g., documents, PDFs), it retrieves relevant content using vector-based retrieval.
   • Web Search: For real-time or broad contextual information, it leverages a web search tool.
   • Recommendation Systems: For personalized queries, it taps into recommendation engines.
3. Data Integration and LLM Synthesis: The retrieved data from the chosen sources is passed to a Large Language Model (LLM). The LLM synthesizes this information into a coherent and contextually relevant response.
4. Output Generation: The system delivers a comprehensive, user-facing answer, which may include references or citations.

Key Features and Advantages

• Centralized Simplicity: Easier to design, implement, and maintain due to a single agent handling all tasks.
• Efficiency & Resource Optimization: Demands fewer computational resources and processes queries quickly due to simpler coordination.
• Dynamic Routing: The agent evaluates each query in real-time to select the most appropriate knowledge source.
• Versatility Across Tools: Supports various data sources and external APIs for both structured and unstructured workflows.
• Ideal for Simpler Systems: Suitable for applications with well-defined tasks or limited integration requirements.

Use Case: Customer Support

Prompt: "Can you tell me the delivery status of my order?" System Process (Single-Agent Workflow):

1. Query Submission and Evaluation: The coordinating agent receives and analyzes the query.
2. Knowledge Source Selection: It retrieves tracking details from an order management database, fetches real-time updates from a shipping provider's API, and optionally conducts a web search for local conditions affecting delivery.
3. Data Integration and LLM Synthesis: The LLM synthesizes this information.
4. Output Generation: The system provides an actionable response. Integrated Response: "Your package is currently in transit and expected to arrive tomorrow evening. The live tracking from UPS indicates it is at the regional distribution center."

4.3.2. Multi-Agent Agentic RAG Systems

Multi-Agent RAG systems represent a modular and scalable evolution, designed to handle complex workflows and diverse query types by leveraging multiple specialized agents. Each agent is optimized for a specific role or data source.

The following figure (Figure 17 from the original paper) shows an overview of multi-agent agentic RAG systems:

Figure 17: An Overview of Multi-Agent Agentic RAG Systems

Workflow

1. Query Submission: A user query is received by a coordinator agent or master retrieval agent. This agent acts as the central orchestrator, delegating the query to specialized retrieval agents.
2. Specialized Retrieval Agents: The query is distributed among multiple retrieval agents, each focusing on a specific type of data source or task:
   • Agent 1: Handles structured queries (e.g., SQL-based databases).
   • Agent 2: Manages semantic searches for unstructured data (e.g., PDFs, internal records).
   • Agent 3: Focuses on real-time public information from web searches or APIs.
   • Agent 4: Specializes in recommendation systems.
3. Tool Access and Data Retrieval: Each agent routes its portion of the query to appropriate tools or data sources within its domain (e.g., vector search, Text-to-SQL, web search, APIs).
4. Data Integration and LLM Synthesis: Once retrieval is complete, the data from all agents is passed to an LLM. The LLM synthesizes the retrieved information into a coherent and contextually relevant response.
5. Output Generation: A comprehensive response is delivered to the user.

Key Features and Advantages

• Modularity: Each agent operates independently, allowing for flexible addition or removal.
• Specialization: Agents can be highly optimized for specific tasks, leading to improved accuracy and retrieval relevance.
• Efficiency: Distributing tasks minimizes bottlenecks and enhances performance for complex workflows.
• Versatility: Suitable for applications spanning multiple domains.

Challenges

• Coordination Complexity: Managing inter-agent communication and task delegation requires sophisticated orchestration mechanisms.
• Computational Overhead: Parallel processing of multiple agents can increase resource usage.
• Data Integration: Synthesizing outputs from diverse sources into a cohesive response is complex.

Use Case: Economic and Environmental Impact Analysis

Prompt: "What are the economic and environmental impacts of renewable energy adoption in Europe?" System Process (Multi-Agent Workflow):

• Agent 1: Retrieves statistical data from economic databases using SQL-based queries.
• Agent 2: Searches for relevant academic papers using semantic search tools.
• Agent 3: Performs a web search for recent news and policy updates.
• Agent 4: Consults a recommendation system for related reports or expert commentary. Response: "Adopting renewable energy in Europe has led to a 20% reduction in greenhouse gas emissions over the past decade, according to EU policy reports. Economically, renewable energy investments have generated approximately 1.2 million jobs, with significant growth in solar and wind sectors. Recent academic studies also highlight potential trade-offs in grid stability and energy storage costs."

4.3.3. Hierarchical Agentic RAG Systems

Hierarchical Agentic RAG systems employ a structured, multi-tiered approach to information retrieval and processing. This architecture involves agents at different levels of abstraction, enabling strategic decision-making and efficient task delegation.

The following figure (Figure 18 from the original paper) shows an illustration of hierarchical agentic RAG:

Figure 18: An illustration of Hierarchical Agentic RAG

Workflow

1. Query Reception: A user query is received by a top-tier agent, responsible for initial assessment and delegation.
2. Strategic Decision-Making: The top-tier agent evaluates the query's complexity and decides which subordinate agents or data sources to prioritize based on reliability or relevance.
3. Delegation to Subordinate Agents: The top-tier agent assigns tasks to lower-level agents specialized in particular retrieval methods (e.g., SQL databases, web search, proprietary systems). These agents execute their assigned tasks independently.
4. Aggregation and Synthesis: Results from subordinate agents are collected and integrated by the higher-level agent, which synthesizes the information into a coherent response.
5. Response Delivery: The final, synthesized answer is returned to the user.

Key Features and Advantages

• Strategic Prioritization: Top-tier agents can prioritize data sources or tasks based on query complexity, reliability, or context.
• Scalability: Distributing tasks across multiple agent tiers enables handling highly complex or multi-faceted queries.
• Enhanced Decision-Making: Higher-level agents apply strategic oversight to improve overall accuracy and coherence of responses.

Challenges

• Coordination Complexity: Maintaining robust inter-agent communication across multiple levels can increase orchestration overhead.
• Resource Allocation: Efficiently distributing tasks among tiers to avoid bottlenecks is non-trivial.

Use Case: Financial Analysis System

Prompt: "What are the best investment options given the current market trends in renewable energy?" System Process (Hierarchical Agentic Workflow):

1. Top-Tier Agent: Assesses query complexity and prioritizes reliable financial databases and economic indicators.
2. Mid-Level Agent: Retrieves real-time market data (e.g., stock prices) from proprietary APIs and structured SQL databases.
3. Lower-Level Agent(s): Conducts web searches for recent policy announcements and consults recommendation systems for expert opinions.
4. Aggregation and Synthesis: The top-tier agent compiles results, integrating quantitative data with policy insights. Response: "Based on current market data, renewable energy stocks have shown a 15% growth over the past quarter, driven by supportive government policies and heightened investor interest. Analysts suggest that wind and solar sectors, in particular, may experience continued momentum, while emerging technologies like green hydrogen present moderate risk but potentially high returns."

4.3.4. Agentic Corrective RAG

Agentic Corrective RAG (CRAG) ensures iterative refinement of context documents and responses, minimizing errors and maximizing relevance through dynamic evaluation and adjustment.

The following figure (Figure 19 from the original paper) shows an overview of agentic corrective RAG:

Figure 19: Overview of Agentic Corrective RAG

Key Ideas of CRAG

CRAG dynamically evaluates and corrects retrieved context to enhance quality. RAG adjusts its approach as follows:

• Document Relevance Evaluation: Retrieved documents are assessed for relevance by a Relevance Evaluation Agent. Documents below a relevance threshold trigger corrective steps.
• Query Refinement and Augmentation: Queries are refined by a Query Refinement Agent, leveraging semantic understanding to optimize retrieval.
• Dynamic Retrieval from External Sources: If context is insufficient, an External Knowledge Retrieval Agent performs web searches or accesses alternative data sources.
• Response Synthesis: Validated and refined information is passed to a Response Synthesis Agent for final generation.

Workflow

The Corrective RAG system is built on five key agents:

1. Context Retrieval Agent: Responsible for retrieving initial context documents from a vector database.
2. Relevance Evaluation Agent: Assesses retrieved documents for relevance and flags irrelevant or ambiguous ones for corrective actions.
3. Query Refinement Agent: Rewrites queries to improve specificity and relevance, using semantic understanding.
4. External Knowledge Retrieval Agent: Performs web searches or accesses alternative data sources when initial context is insufficient.
5. Response Synthesis Agent: Synthesizes all validated information into a coherent and accurate response.

Key Features and Advantages

• Iterative Correction: Ensures high response accuracy by dynamically identifying and correcting irrelevant or ambiguous retrieval results.
• Dynamic Adaptability: Incorporates real-time web searches and query refinement.
• Agentic Modularity: Each agent performs specialized tasks, ensuring efficient and scalable operation.
• Factuality Assurance: Validating all retrieved and generated content minimizes the risk of hallucination or misinformation.

Use Case: Generative AI Research Query

Prompt: "What are the latest findings in generative AI research?" System Process (Corrective RAG Workflow):

1. Query Submission: User submits the query.
2. Context Retrieval: Context Retrieval Agent retrieves initial documents from a database of published papers.
3. Relevance Evaluation: Relevance Evaluation Agent assesses document alignment with the query, classifying them as relevant, ambiguous, or irrelevant, flagging irrelevant ones for correction.
4. Corrective Actions (if needed): Query Refinement Agent rewrites the query. External Knowledge Retrieval Agent performs web searches to fetch additional papers and reports.
5. Response Synthesis: Response Synthesis Agent integrates validated documents into a summary. Response: "Recent findings in generative AI highlight advancements in diffusion models, reinforcement learning for text-to-video tasks, and optimization techniques for large-scale model training. For more details, refer to studies published in NeurIPS 2024 and AAAI 2025."

4.3.5. Adaptive Agentic RAG

Adaptive Retrieval-Augmented Generation (Adaptive RAG) enhances flexibility and efficiency by dynamically adjusting retrieval strategies based on query complexity. It may even bypass retrieval for straightforward queries.

The following figure (Figure 20 from the original paper) shows an overview of adaptive agentic RAG:

Figure 20: An Overview of Adaptive Agentic RAG

Key Ideas of Adaptive RAG

The critical innovation is the dynamic adjustment of RAG strategies based on query complexity.

• Straightforward Queries: For simple fact-based questions, the system directly generates an answer using pre-existing LLM knowledge, avoiding retrieval.
• Simple Queries: For moderately complex tasks requiring minimal context, the system performs a single-step retrieval.
• Complex Queries: For multi-layered queries requiring iterative reasoning, the system employs multi-step retrieval, progressively refining intermediate results.

Workflow

The Adaptive RAG system is built on three primary components:

1. Classifier Role: A smaller language model analyzes the query to predict its complexity, trained on automatically labeled datasets from past model outcomes.
2. Dynamic Strategy Selection:
   • For straightforward queries, it avoids retrieval.
   • For simple queries, it uses single-step retrieval.
   • For complex queries, it employs multi-step retrieval.
3. LLM Integration: The LLM synthesizes retrieved information. Iterative interactions between the LLM and the classifier enable refinement.

Key Features and Advantages

• Dynamic Adaptability: Adjusts retrieval strategies based on query complexity, optimizing computational efficiency and response accuracy.
• Resource Efficiency: Minimizes unnecessary overhead for simple queries while ensuring thorough processing for complex ones.
• Enhanced Accuracy: Iterative refinement ensures complex queries are resolved with high precision.
• Flexibility: Can be extended to incorporate additional pathways like domain-specific tools.

Use Case: Package Delay Inquiry

Prompt: "Why is my package delayed, and what alternatives do I have?" System Process (Adaptive RAG Workflow):

1. Query Classification: The system classifies the query as complex due to its multi-faceted nature (reason for delay + alternatives).
2. Multi-Step Retrieval:
   • Retrieves tracking details from the order database.
   • Fetches real-time status updates from the shipping provider API.
   • Conducts a web search for external factors (e.g., weather conditions).
3. Response Synthesis: The LLM integrates all retrieved information. Response: "Your package is delayed due to severe weather conditions in your region. It's currently at the local distribution center and will be delivered tomorrow. Alternatively, you may pick up your package from the facility."

4.3.6. Graph-Based Agentic RAG

This category integrates graph-based knowledge structures with agentic retrieval to enhance reasoning and retrieval accuracy, especially for tasks requiring relational understanding.

4.3.6.1. Agent-G: Agentic Framework for Graph RAG

Agent-G is an agentic framework that combines structured and unstructured data sources for Retrieval-Augmented Generation (RAG), improving reasoning and retrieval accuracy. It utilizes modular retriever banks, dynamic agent interaction, and feedback loops.

The following figure (Figure 21 from the original paper) shows an overview of Agent-G: Agentic Framework for Graph RAG:

Figure 21: An Overview of Agent-G: Agentic Framework for Graph RAG [8]

Key Idea of Agent-G

Agent-G's core principle is to dynamically assign retrieval tasks to specialized components:

• Graph Knowledge Bases: Utilizes structured data to extract relationships, hierarchies, and connections (e.g., disease-to-symptom mappings).
• Unstructured Documents: Traditional text retrieval systems provide contextual information to complement graph data.
• Critic Module: Evaluates the relevance and quality of retrieved information, ensuring alignment with the query.
• Feedback Loops: Refines retrieval and synthesis through iterative validation and re-querying.

Workflow

The Agent-G system is built on four primary components:

1. Retriever Bank: A modular set of agents specializes in retrieving graph-based or unstructured data, dynamically selecting relevant sources.
2. Critic Module: Validates retrieved data for relevance and quality, flagging low-confidence results for re-retrieval.
3. Dynamic Agent Interaction: Task-specific agents collaborate to integrate diverse data types, ensuring cohesive retrieval and synthesis.
4. LLM Integration: Synthesizes validated data into a coherent response, with iterative feedback from the critic.

Key Features and Advantages

• Enhanced Reasoning: Combines structured relationships from graphs with contextual information from unstructured documents.
• Dynamic Adaptability: Adjusts retrieval strategies dynamically based on query requirements.
• Improved Accuracy: Critic module reduces the risk of irrelevant or low-quality data.
• Scalable Modularity: Supports the addition of new agents for specialized tasks.

Use Case: Medical Knowledge Query

Prompt: "What are the common symptoms of Type 2 Diabetes, and how are they related to heart disease?" System Process (Agent-G Workflow):

1. Query Reception and Assignment: System identifies need for both graph-structured and unstructured data.
2. Graph Retriever: Extracts relationships between Type 2 Diabetes and heart disease from a medical knowledge graph, identifying shared risk factors.
3. Document Retriever: Retrieves Type 2 Diabetes symptoms from medical literature, adding contextual information.
4. Critic Module: Evaluates relevance and quality of retrieved graph data and document data, flagging low-confidence results.
5. Response Synthesis: The LLM integrates validated data from both retrievers. Response: "Type 2 Diabetes symptoms include increased thirst, frequent urination, and fatigue. Studies show a 50% correlation between diabetes and heart disease, primarily through shared risk factors such as obesity and high blood pressure."

4.3.6.2. GeAR: Graph-Enhanced Agent for Retrieval-Augmented Generation

GeAR introduces an agentic framework that enhances traditional RAG systems by incorporating graph-based retrieval mechanisms. It leverages graph expansion techniques and an agent-based architecture to improve multi-hop retrieval and handle complex queries.

The following figure (Figure 22 from the original paper) shows an overview of GeAR: Graph-Enhanced Agent for Retrieval-Augmented Generation:

Figure 22: An Overview of GeAR: Graph-Enhanced Agent for Retrieval-Augmented Generation[35]

Key Idea of GeAR

GeAR advances RAG performance through two primary innovations:

• Graph Expansion: Enhances conventional base retrievers (e.g., BM25) by expanding retrieval to include graph-structured data, capturing complex relationships between entities. It identifies and retrieves directly connected entities.
• Agent Framework: Incorporates an agent-based architecture to manage retrieval tasks more effectively, allowing for dynamic and autonomous decision-making in the retrieval process.

Workflow

The GeAR system operates through the following components:

1. Graph Expansion Module: Integrates graph-based data into the retrieval process, considering relationships between entities. It identifies and retrieves directly connected entities.
2. Agent-Based Retrieval: Employs an agent framework to manage retrieval, enabling dynamic selection and combination of strategies. Agents autonomously decide to utilize graph-expanded retrieval paths.
3. LLM Integration: Combines the retrieved information, enriched by graph expansion, with the LLM's capabilities to generate coherent and contextually relevant responses.

Key Features and Advantages

• Enhanced Multi-Hop Retrieval: Graph expansion allows reasoning over multiple interconnected pieces of information.
• Agentic Decision-Making: Agent framework enables dynamic and autonomous selection of retrieval strategies.
• Improved Accuracy: Incorporating structured graph data enhances the precision of retrieved information.
• Scalability: Modular nature allows integration of additional retrieval strategies and data sources.

Use Case: Literary Influence Query

Prompt: "Which author influenced the mentor of J.K. Rowling?" System Process (GeAR Workflow):

1. Top-Tier Agent: Evaluates the query's multi-hop nature and determines the need for graph expansion and document retrieval.
2. Graph Expansion Module: Identifies J.K. Rowling's mentor as a key entity and traces literary influences on that mentor through graph-structured data.
3. Agent-Based Retrieval: An agent autonomously selects the graph-expanded retrieval path and integrates additional context from textual data sources.
4. Response Synthesis: Combines insights from graph and document retrieval using the LLM. Response: "J.K. Rowling's mentor, [Mentor Name], was heavily influenced by [Author Name], known for their notable work in [Genre]. This highlights the intricate relationships in literary circles, where influential ideas often pass through multiple generations of authors."

4.3.7. Agentic Document Workflows (ADW)

Agentic Document Workflows (ADW) extend traditional RAG paradigms by enabling end-to-end knowledge work automation specifically focused on document processing. ADW combines Intelligent Document Processing (IDP) with RAG through agentic orchestration, multi-step workflows, and domain-specific logic.

The following figure (Figure 23 from the original paper) shows an overview of agentic document workflows (ADW):

Figure 23: An Overview of Agentic Document Workflows (ADW) [36]

Workflow

1. Document Parsing and Information Structuring: Documents are parsed using enterprise-grade tools (e.g., LlamaParse) to extract relevant data fields (e.g., invoice numbers, dates). Structured data is organized for downstream processing.
2. State Maintenance Across Processes: The system maintains state about document context, ensuring consistency and relevance across multi-step workflows, tracking document progression.
3. Knowledge Retrieval: Relevant references are retrieved from external knowledge bases (e.g., LlamaCloud) or vector indexes. Real-time, domain-specific guidelines are retrieved for enhanced decision-making.
4. Agentic Orchestration: Intelligent agents apply business rules, perform multi-hop reasoning, and generate actionable recommendations. It orchestrates components like parsers, retrievers, and external APIs.
5. Actionable Output Generation: Outputs are presented in structured formats, tailored to specific use cases, with recommendations and extracted insights synthesized into concise reports.

Key Features and Advantages

• State Maintenance: Tracks document context and workflow stage, ensuring consistency.
• Multi-Step Orchestration: Handles complex workflows involving multiple components and external tools.
• Domain-Specific Intelligence: Applies tailored business rules and guidelines for precise recommendations.
• Scalability: Supports large-scale document processing with modular and dynamic agent integration.
• Enhanced Productivity: Automates repetitive tasks while augmenting human expertise.

Use Case: Invoice Payments Workflow

Prompt: "Generate a payment recommendation report based on the submitted invoice and associated vendor contract terms." System Process (ADW Workflow):

1. Parse the invoice to extract key details (invoice number, date, vendor, line items, payment terms).
2. Retrieve the corresponding vendor contract to verify payment terms, applicable discounts, or compliance requirements.
3. Generate a payment recommendation report, including original amount due, potential early payment discounts, budget impact analysis, and strategic payment actions. Response: "Invoice INV-2025-045 for 15,000.00 has been processed. An early payment discount of 2% is available if paid by 2025-04-10, reducing the amount due to14,700.00. A bulk order discount of 5% was applied as the subtotal exceeded $10,000.00. It is recommended to approve early payment to save 2% and ensure timely fund allocation for upcoming project phases."

5. Experimental Setup

This paper is a survey, and as such, it does not present new experimental results from its own research. Instead, it provides an overview of the landscape of Agentic RAG systems, including tools, frameworks, and benchmarks used to evaluate RAG systems in general, which would apply to Agentic RAG as well. Therefore, this section will describe the datasets and benchmarks typically used for evaluating RAG and agent-based systems, as outlined in the paper's Section 9.

5.1. Datasets

The paper lists various datasets relevant for evaluating RAG systems across different downstream tasks. These datasets are chosen to validate the effectiveness of retrieval and generation components, often focusing on question answering, dialog, and reasoning.

The following are the results from Table 3 of the original paper:

Some key datasets and their characteristics:

• Natural Questions (NQ) [65]: A large-scale QA dataset for open-domain question answering, where questions are naturally occurring Google queries and answers are derived from Wikipedia articles. It includes both short and long answers.

• SQuAD (Stanford Question Answering Dataset) [67]: A reading comprehension dataset, where questions are based on Wikipedia articles, and the answer to every question is a segment of text from the corresponding reading passage.

• HotpotQA [60]: A multi-hop QA dataset that requires reasoning over multiple documents to answer questions, often involving finding and combining information from different sources.

• MS MARCO (Microsoft Machine Reading Comprehension) [56]: Focuses on passage ranking and question answering, widely used for dense retrieval tasks. Questions are anonymized Bing queries.

• WikiText-103 [99]: A large language modeling dataset composed of over 103 million words extracted from Wikipedia articles, often used to evaluate the perplexity and generation quality of LLMs.

• FEVER (Fact Extraction and VERification) [90]: A dataset for fact checking and verification, requiring systems to determine the veracity of claims by retrieving and evaluating evidence from Wikipedia.

  These datasets are chosen because they represent diverse challenges for RAG systems, ranging from simple fact retrieval to complex multi-hop reasoning, long-form generation, and domain-specific applications. They are effective for validating a method's ability to retrieve relevant information and generate accurate, coherent responses.

5.2. Evaluation Metrics

As a survey paper, this work does not propose or use new evaluation metrics. However, the benchmarks and datasets listed in Section 9 of the paper imply the use of standard metrics commonly employed to evaluate Retrieval-Augmented Generation (RAG) systems and their components. These metrics typically assess both the retrieval quality and the generation quality.

Common metrics for evaluating RAG systems, often used in conjunction with the listed benchmarks, include:

5.2.1. Retrieval Metrics

These metrics assess how well the system identifies and retrieves relevant documents or passages.

• Precision ($P$): The proportion of retrieved documents that are relevant. $ P = \frac{\text{Number of relevant documents retrieved}}{\text{Total number of documents retrieved}} $ Where:
  • Number of relevant documents retrieved is the count of items that are both relevant to the query and were returned by the system.
  • Total number of documents retrieved is the total count of items returned by the system.
• Recall ($R$): The proportion of all relevant documents in the corpus that were retrieved by the system. $ R = \frac{\text{Number of relevant documents retrieved}}{\text{Total number of relevant documents in the corpus}} $ Where:
  • Number of relevant documents retrieved is the count of items that are both relevant to the query and were returned by the system.
  • Total number of relevant documents in the corpus is the total count of items in the entire collection that are relevant to the query.
• F1-Score: The harmonic mean of Precision and Recall, providing a single score that balances both. $ F1 = 2 \times \frac{P \times R}{P + R} $ Where:
  • $P$ is Precision.
  • $R$ is Recall.
• Mean Reciprocal Rank (MRR): For ranked lists of results, MRR measures the average of the reciprocal ranks of the first relevant document for a set of queries. $ MRR = \frac{1}{|Q|} \sum_{i=1}^{|Q|} \frac{1}{\text{rank}_i} $ Where:
  • $|Q|$ is the total number of queries.
  • $\text{rank}_i$ is the rank of the first relevant document for the $i$-th query. If no relevant document is found, the reciprocal rank is 0.
• Normalized Discounted Cumulative Gain (NDCG): Measures the ranking quality, considering the position of relevant documents and their relevance scores. Highly relevant documents appearing early in the results list increase NDCG. $ NDCG_k = \frac{DCG_k}{IDCG_k} $ Where:
  • $DCG_k = \sum_{i=1}^{k} \frac{rel_i}{\log_2(i+1)}$ is the Discounted Cumulative Gain at rank $k$.
  • $IDCG_k$ is the Ideal Discounted Cumulative Gain at rank $k$ (the maximum possible $DCG_k$ given the relevant documents).
  • $rel_i$ is the relevance score of the document at rank $i$.

5.2.2. Generation Metrics

These metrics evaluate the quality, coherence, and factual accuracy of the text generated by the LLM.

• ROUGE (Recall-Oriented Understudy for Gisting Evaluation): A set of metrics used for evaluating automatic summarization and machine translation. It works by comparing an automatically produced summary or translation with a set of reference summaries (human-produced).
  • ROUGE-N: Measures the overlap of N-grams (contiguous sequences of N items) between the system-generated text and reference texts. $ \text{ROUGE-N} = \frac{\sum_{\text{sentence} \in \text{ReferenceSummaries}} \sum_{n\text{-gram} \in \text{sentence}} \text{Count}{\text{match}}(n\text{-gram})}{\sum{\text{sentence} \in \text{ReferenceSummaries}} \sum_{n\text{-gram} \in \text{sentence}} \text{Count}(n\text{-gram})} $ Where:
    • n-gram refers to a contiguous sequence of $n$ words.
    • Count_match(n-gram) is the maximum number of n-grams co-occurring in the system summary and a reference summary.
    • Count(n-gram) is the number of n-grams in the reference summary.
  • ROUGE-L: Based on the Longest Common Subsequence (LCS), which accounts for sentence-level structure.
• BLEU (Bilingual Evaluation Understudy): A metric for evaluating the quality of text which has been machine-translated from one natural language to another. It compares overlapping n-grams with reference translations.
• METEOR (Metric for Evaluation of Translation with Explicit Ordering): A metric for the evaluation of machine translation output that addresses some shortcomings of BLEU by including factors like stemming and synonymy.
• Perplexity (PPL): A measure of how well a probability model predicts a sample. In LLMs, a lower perplexity generally indicates a better model. $ PPL(W) = \sqrt[N]{\prod_{i=1}^{N} \frac{1}{P(w_i | w_1, \dots, w_{i-1})}} $ Where:
  • $W = (w_1, w_2, \dots, w_N)$ is a sequence of $N$ words.
  • $P(w_i | w_1, \dots, w_{i-1})$ is the probability of the $i$-th word given the preceding words, as predicted by the language model.
• Factual Accuracy: Often measured qualitatively through human evaluation or quantitatively using specialized fact-checking datasets (like FEVER) or QA datasets where answers can be directly verified.
• Hallucination Rate: The frequency at which the LLM generates factually incorrect information that is not supported by the retrieved context. This is typically assessed through human evaluation.

5.3. Baselines

The paper, being a survey, does not present its own baselines. However, the discussion of the evolution of RAG paradigms (Naïve RAG, Advanced RAG, Modular RAG, Graph RAG) implicitly positions these as foundational baselines against which Agentic RAG systems demonstrate their advancements. When individual Agentic RAG systems (like Agent-G or GeAR) are proposed in research, they are typically compared against:

• Traditional RAG variants: Naive RAG or Advanced RAG often serve as a starting point to show the benefits of adding agentic capabilities.

• State-of-the-art RAG models: More sophisticated RAG architectures that do not incorporate full agentic intelligence.

• Pure LLM generation: To highlight the benefit of retrieval augmentation itself.

• Other agentic or multi-agent systems: Especially if the focus is on the agentic orchestration aspect rather than just RAG.

  The specific benchmarks mentioned (e.g., BEIR, MS MARCO, TREC, HotpotQA) are designed to facilitate such comparisons by providing standardized tasks and evaluation methodologies. For instance, AgentG is explicitly mentioned as a benchmark tailored for agentic RAG tasks, implying it would be used to compare different agentic RAG frameworks.

6. Results & Analysis

As a survey paper, this document does not present novel experimental results or quantitative data from the authors' own research. Instead, it systematically categorizes and analyzes existing Agentic RAG systems, highlighting their mechanisms, strengths, weaknesses, and appropriate use cases. The "results" section of this analysis will focus on the comparative analysis provided by the paper, which serves to differentiate Agentic RAG from its predecessors and highlight the specific benefits and challenges of various Agentic RAG architectures.

6.1. Core Results Analysis

The paper's core analysis is presented through a comparative table that positions Agentic RAG and Agentic Document Workflows (ADW) against Traditional RAG. This comparison emphasizes the evolutionary advancements and the paradigm shift introduced by agentic approaches.

The following are the results from Table 2 of the original paper:

6.1.1. Comparison with Traditional RAG

• Focus: Traditional RAG is limited to isolated retrieval and generation. In contrast, Agentic RAG emphasizes multi-agent collaboration and reasoning, while ADW focuses on document-centric end-to-end workflows. This highlights a shift from component-level enhancement to system-level intelligence and automation.
• Context Maintenance: Traditional RAG has limited context maintenance, often struggling to remember previous interactions or refine understanding over time. Agentic RAG significantly improves this through dedicated memory modules for both short-term and long-term context. ADW further specializes this by maintaining state across multi-step workflows specifically for documents.
• Dynamic Adaptability: Traditional RAG offers minimal dynamic adaptability due to its static nature. Agentic RAG boasts high adaptability, able to adjust strategies in real-time. ADW provides adaptability tailored to document workflows. This is a critical advantage for handling complex, evolving queries.
• Workflow Orchestration: Workflow orchestration is largely absent in Traditional RAG. Agentic RAG explicitly orchestrates multi-agent tasks, and ADW integrates multi-step document processing, demonstrating advanced control over task execution.
• External Tool Use: Traditional RAG has basic tool integration (e.g., for retrieval). Agentic RAG extends this significantly, using APIs and knowledge bases more broadly. ADW shows the deepest integration, leveraging business rules and domain-specific tools. This indicates a progression towards more capable and interconnected AI systems.
• Scalability: Traditional RAG is limited to smaller datasets or queries. Agentic RAG is presented as scalable for multi-agent systems, and ADW for multi-domain enterprise workflows. This suggests agentic approaches are better equipped for real-world, large-scale deployments.
• Complex Reasoning: Traditional RAG offers basic reasoning (simple Q&A). Agentic RAG provides multi-step reasoning with agents, while ADW offers structured reasoning across documents. This is a major leap in handling intricate problem-solving.
• Strengths: Traditional RAG excels in simplicity and quick setup. Agentic RAG offers high accuracy and collaborative reasoning. ADW provides end-to-end automation and domain-specific intelligence.
• Challenges: Traditional RAG struggles with poor contextual understanding. Agentic RAG faces coordination complexity among agents. ADW contends with resource overhead and domain standardization. These highlight that while agentic systems are powerful, they introduce new complexities in management and resource intensity.

6.1.2. Applications of Agentic RAG

The paper details various applications across industries, validating Agentic RAG's versatility and effectiveness in real-world scenarios.

• Customer Support and Virtual Assistants: Improves response quality and operational efficiency by providing personalized, context-aware replies and real-time adaptability. Example: Twitch's ad sales enhancement using Agentic RAG on Amazon Bedrock.

• Healthcare and Personalized Medicine: Enables personalized care and time efficiency by retrieving real-time clinical guidelines and patient history for diagnostics and treatment. Example: Patient case summary generation.

• Legal and Contract Analysis: Enhances risk identification and efficiency in legal workflows. Example: Automated contract review to flag deviations and ensure compliance.

• Finance and Risk Analysis: Provides real-time analytics and risk mitigation for investment decisions and market analysis. Example: Auto insurance claims processing, generating recommendations with regulatory compliance.

• Education and Personalized Learning: Facilitates tailored learning paths and engaging interactions. Example: Research paper generation for higher education, synthesizing findings and providing summaries.

• Graph-Enhanced Applications in Multimodal Workflows: Combines graph structures with retrieval for multi-modal capabilities (text, images, video). Example: Market survey generation for product trends, enriching reports with multimedia.

  These applications collectively demonstrate that Agentic RAG is not just a theoretical improvement but a practical solution driving AI innovation in diverse, complex domains. The examples underscore its ability to handle dynamic adaptability, contextual precision, and knowledge-intensive challenges.

6.2. Ablation Studies / Parameter Analysis

The survey paper does not include ablation studies or parameter analysis, as it is a review of existing work rather than new experimental research. Such analyses would typically be found in individual research papers proposing specific Agentic RAG models or architectures, where components are systematically removed or parameters varied to understand their impact on performance.

7. Conclusion & Reflections

7.1. Conclusion Summary

The paper concludes that Agentic Retrieval-Augmented Generation (RAG) represents a transformative advancement in artificial intelligence. It successfully addresses the limitations of traditional RAG systems by integrating autonomous agents, which leverage agentic design patterns like reflection, planning, tool use, and multi-agent collaboration. This integration enables Agentic RAG systems to tackle complex, real-world tasks with enhanced precision and adaptability.

The survey meticulously traces the evolution of RAG paradigms, from Naïve RAG to Modular and Graph RAG, highlighting how Agentic RAG has emerged as a pivotal development. It overcomes static workflows and limited contextual adaptability, delivering unparalleled flexibility, scalability, and context-awareness. The paper provides a comprehensive taxonomy of Agentic RAG architectures (e.g., Single-Agent, Multi-Agent, Hierarchical, Corrective, Adaptive, Graph-Based, and Agentic Document Workflows) and showcases its broad applicability across critical sectors such as healthcare, finance, education, and creative industries. Furthermore, it discusses practical implementation tools and frameworks, solidifying Agentic RAG's role as a cornerstone for next-generation AI applications.

7.2. Limitations & Future Work

The authors acknowledge that despite its promise, Agentic RAG systems face significant challenges that require ongoing research.

• Coordination Complexity: Managing interactions between multiple autonomous agents is inherently complex, leading to orchestration overhead.

• Computational Overhead: The dynamic and multi-agent nature of these systems can significantly increase resource requirements, especially for high query volumes or complex workflows.

• Scalability Limitations: While generally scalable for multi-agent systems, the dynamic nature can strain computational resources under high demand.

• Ensuring Ethical Decision-Making: The autonomy of agents introduces challenges in ensuring ethical decision-making and preventing unintended biases or harmful actions.

• Performance Optimization: Optimizing performance for real-world applications remains a challenge, requiring careful tuning and robust engineering.

• Explainability: The black-box nature of LLMs combined with complex agent interactions can make it difficult to understand why an Agentic RAG system arrived at a particular conclusion, posing challenges for trust and auditing.

  The paper suggests future research directions should focus on:

• Addressing the aforementioned challenges, particularly coordination complexity and computational efficiency.

• Developing more robust mechanisms for ethical decision-making and bias mitigation in autonomous agents.

• Enhancing the explainability and interpretability of Agentic RAG systems.

• Further exploring the unique aspects of Agentic RAG, such as multi-agent collaboration and dynamic adaptability, to advance the field.

• Developing new benchmarks and evaluation metrics that specifically capture the performance of agentic and dynamic RAG systems, moving beyond traditional RAG evaluation.

7.3. Personal Insights & Critique

This survey provides an exceptionally well-structured and comprehensive overview of Agentic RAG, effectively positioning it as the next frontier in LLM capabilities. The detailed taxonomy and workflow patterns are particularly valuable, offering a clear mental model for understanding the diverse approaches within this emerging field. The inclusion of practical examples for each Agentic RAG type significantly aids in grasping their real-world applicability. For a novice, the progressive explanation of RAG paradigms from Naïve to Agentic is highly effective in building foundational knowledge.

Inspirations and Applications to Other Domains: The core idea of embedding autonomous agents within an information retrieval and generation pipeline is highly transferable.

• Scientific Discovery: An Agentic RAG system could orchestrate agents to scour scientific literature, run simulations via external tools, analyze experimental data, and propose new hypotheses, accelerating research in fields like material science or drug discovery.
• Personalized Legal Assistant for Laypersons: Imagine an agent that not only retrieves legal statutes but can also plan steps for a legal process, reflect on potential outcomes, and use tools to draft preliminary documents or identify relevant precedents, guiding individuals through complex legal challenges.
• Dynamic Educational Content Generation: Beyond simple question answering, an Agentic RAG system could dynamically generate personalized learning modules, adapt to a student's learning style, create interactive exercises, and even collaborate with a "teacher agent" for feedback and assessment.

Potential Issues, Unverified Assumptions, or Areas for Improvement:

1. Orchestration Overhead and Debuggability: While the paper acknowledges coordination complexity as a challenge, the practical difficulties of debugging, monitoring, and maintaining multi-agent systems in production are immense. Failures can cascade, and identifying the root cause in a dynamic, collaborative system is far harder than in a sequential pipeline. The cognitive load on engineers to manage these systems will be substantial.

2. Cost and Latency: The "computational overhead" mentioned is a significant practical barrier. Each agent, especially when powered by LLMs, can incur substantial costs and introduce latency. For real-time applications, balancing agentic intelligence with efficiency remains a critical engineering challenge. The paper notes that Adaptive RAG tries to optimize this by avoiding retrieval for simple queries, but this is a specific solution, not a general one for all agentic types.

3. Ethical Risks and Guardrails: The increased autonomy of Agentic RAG amplifies ethical risks (e.g., bias amplification, misinformation propagation, unintended actions). While ethical decision-making is listed as a challenge, the mechanisms for implementing robust guardrails, human-in-the-loop interventions, and accountability frameworks in such complex, dynamic systems need much deeper exploration.

4. Evaluation of Agentic Behavior: The listed benchmarks are primarily for RAG output quality. Evaluating the agentic behaviors (planning, reflection, tool use, collaboration) themselves, beyond just the final output, is an nascent research area. How do we rigorously measure the "quality" of a plan or the "effectiveness" of reflection? This will require new types of benchmarks and metrics.

5. Standardization and Interoperability: The proliferation of agentic frameworks and tools (e.g., LangChain, LlamaIndex, CrewAI, AutoGen) hints at a lack of standardization. As Agentic RAG matures, interoperability standards will be crucial for building complex, robust systems from diverse components.

   Overall, this survey is an excellent resource for understanding the current state and future trajectory of RAG. It clearly articulates the shift towards more intelligent and adaptive LLM systems, laying the groundwork for future research and development in this exciting area. The challenges highlighted are not deterrents but clear indicators of where the next wave of innovation in AI will focus.