SoK: Agentic RAG — Taxonomy, Architectures, Evaluation (2026)

Free

First Systematization of Knowledge for Agentic RAG — formalizes retrieval-generation loops as finite-horizon POMDPs; multi-dimensional taxonomy covering planning strategies, retrieval orchestration, memory paradigms, and tool coordination

FreeFree tier

Type

Open Source

Visit Website GitHub (40 stars)

About SoK: Agentic RAG — Taxonomy, Architectures, Evaluation (2026)

This Systematization of Knowledge (SoK) paper provides the first unified framework for understanding Agentic Retrieval-Augmented Generation (RAG) systems as sequential decision-making processes. It formalizes agentic retrieval-generation loops as finite-horizon partially observable Markov decision processes (POMDPs), explicitly modeling control policies and state transitions. The paper develops a comprehensive taxonomy and modular architectural decomposition that categorizes systems by planning mechanisms, retrieval orchestration, memory paradigms, and tool-invocation behaviors. It also analyzes critical limitations of traditional static evaluation practices, identifies severe systemic risks such as compounding hallucination propagation, memory poisoning, retrieval misalignment, and cascading tool-execution vulnerabilities, and outlines key research directions for building reliable, controllable, and scalable agentic retrieval systems.

Key Features

Formalizes agentic retrieval-generation loops as finite-horizon POMDPs with explicit control policies and state transitions

Develops a comprehensive taxonomy and modular architectural decomposition covering planning, retrieval orchestration, memory paradigms, and tool invocation

Analyzes limitations of traditional static evaluation practices for autonomous RAG systems

Identifies systemic risks: compounding hallucination propagation, memory poisoning, retrieval misalignment, and cascading tool-execution vulnerabilities

Outlines doctoral-scale research directions: stable adaptive retrieval, cost-aware orchestration, formal trajectory evaluation, and oversight mechanisms

Pros & Cons

Pros

First systematic and unified framework for Agentic RAG, bridging fragmented research
Rigorous formalization using POMDPs provides theoretical foundation for sequential decision-making
Comprehensive taxonomy enables clear categorization and comparison of different agentic RAG approaches
Explicitly identifies and analyzes critical risks like hallucination propagation and memory poisoning that are often overlooked
Provides actionable research roadmap for building more reliable agentic retrieval systems

Cons

Primarily a theoretical framework and survey; does not include practical implementations or code
Does not provide empirical evaluation of the proposed taxonomy or formalization
Limited to analysis of existing literature; may not cover all emerging approaches
As a research paper, it is not a directly usable software tool or library

Best For

Understanding and classifying agentic RAG architectures for research and developmentGuiding the design of reliable, controllable, and scalable retrieval-augmented generation systemsEvaluating autonomous RAG systems with awareness of static evaluation limitations and systemic risksIdentifying open challenges and research directions for academic theses and industry innovation

FAQ

What is the main contribution of this paper?

The paper provides the first Systematization of Knowledge (SoK) for Agentic RAG. It formalizes agentic retrieval-generation loops as finite-horizon POMDPs, develops a comprehensive taxonomy covering planning, retrieval orchestration, memory paradigms, and tool invocation, and identifies critical risks and research directions.

What are the systemic risks identified for agentic RAG?

The paper identifies compounding hallucination propagation, memory poisoning, retrieval misalignment, and cascading tool-execution vulnerabilities as severe risks inherent to autonomous loops.

What research directions does the paper propose?

It outlines research directions including stable adaptive retrieval, cost-aware orchestration, formal trajectory evaluation, and oversight mechanisms for building reliable, controllable, and scalable agentic retrieval systems.