RAG 2.0 (GraphRAG/Agentic/CRAG)

Source: RAG 2.0 기술 총정리 — from concepts to practical application

While RAG 1.0 focused on retrieval, RAG 2.0 focuses on verification. Three distinct approaches each tackle the verification problem in their own way.

Limitations of Traditional RAG (Naive RAG)

Classic RAG is a three-stage pipeline: document indexing → vector similarity search → LLM generation.

Limitation	Description
Retrieval quality	Vector similarity alone cannot judge true contextual relevance
Global questions	Cannot answer questions like “What are the main themes in this dataset?”
Connecting relationships	Struggles with complex questions that require linking multiple documents
No self-verification	Cannot judge on its own whether the retrieved documents are actually relevant

The Three Approaches of RAG 2.0

    flowchart LR
    A["RAG 1.0\nVector Search"] --> B["GraphRAG\nKnowledge Structure Verification"]
    A --> C["Agentic RAG\nRetrieval Strategy Verification"]
    A --> D["CRAG\nInformation Purity Verification"]

1. GraphRAG — Knowledge-Graph-Based Retrieval

Announced and open-sourced by Microsoft in 2024, this approach indexes documents as a knowledge graph rather than as vectors.

Indexing Stage

Split documents into TextUnits
Use an LLM to extract entities, relationships, and claims
Structure the knowledge graph
Cluster communities (tightly connected groups of entities) using graph algorithms
Generate an LLM summary for each community

Query Stage — Two Retrieval Methods

Method	Target questions	Example
Local search	Related to a specific entity	“What projects relate to policy issues commissioned by the City of Seoul?”
Global search	Related to the entire dataset	“What are the main trends across our department’s bidding projects?”

Performance and Practical Considerations

Comprehensiveness: 70-80% win rate over traditional RAG (Microsoft benchmark)
High indexing cost: the LLM is called multiple times over every document — watch API costs
Domain tuning required: entity/relationship extraction prompts need to be tuned per domain
GraphRAG 1.0 (Dec 2024): integrates vector stores such as LanceDB and Azure AI Search

2. Agentic RAG — Dynamically Reasoning Retrieval

Integrates an AI agent into the RAG pipeline to dynamically formulate and execute a retrieval strategy.

ReAct Framework Flow

    flowchart TD
    A["Question Input"] --> B["Agent: Analyze Question & Plan"]
    B --> C["Select Tool\n(Vector Search / Web Search / DB Query)"]
    C --> D["Evaluate Results"]
    D -->|"Insufficient Information"| C
    D -->|"Sufficient Information"| E["Generate Final Answer"]

Practical Example — Reviewing a New Bidding Project

Vector DB search → finds 5 similar past projects
Graph search → checks competitors’ contract-award status
Automatic judgment → “Competitor A has won 2 City of Seoul projects; strategy needs to be reconsidered”

Key Characteristics

Multiple data sources: chooses among vector DB, RDB, web search, and APIs as needed
Query routing: automatically connects to the appropriate search engine based on question type
Multi-turn: gathers information incrementally through multiple rounds of interaction

Implementation Tools

Tool	Characteristics
LangGraph	State-graph-based workflow, integrates with LangChain
LlamaIndex	Built-in agentic document workflows
AutoGen	Microsoft’s multi-agent framework
CrewAI	Collaborative agent-team system

3. CRAG — Self-Verifying Retrieval

Published on arXiv in January 2024. A mechanism that evaluates and corrects the quality of retrieved documents on its own.

Three Core Components

① Retrieval Evaluator

Assigns a relevance score to each retrieved document and chooses one of three paths.

    flowchart TD
    A["Search Results"] --> B{"Relevance Evaluation"}
    B -->|"Correct\nHighly Relevant"| C["Refine Document, Then Use"]
    B -->|"Incorrect\nNot Relevant"| D["Fall Back to Web Search"]
    B -->|"Ambiguous\nUncertain"| E["Refine + Web Search in Parallel"]
    C --> F["Final Answer"]
    D --> F
    E --> F

② Knowledge Refinement

Decompose documents into “knowledge strips”
Re-evaluate the relevance of each fragment
Select only the core information

③ Web Search Expansion

Perform web search after query rewriting
Prioritize trustworthy sources such as Wikipedia

The Strength of CRAG — Modularity

Because it only requires adding an evaluator layer to an existing RAG pipeline, it can be adopted incrementally. It is the most practical entry point into RAG 2.0, requiring no full rebuild.

Comparing the Three Techniques

	GraphRAG	Agentic RAG	CRAG
Core innovation	Structuring knowledge as a graph	Dynamic agent	Self-verification
What it verifies	The structure of knowledge	The appropriateness of retrieval	The purity of information
Strength	Complex relationships, global questions	Flexible, multi-source	Simple, modular
Weakness	High cost, slow indexing	Complexity, hard to debug	Depends on evaluator accuracy
Adoption difficulty	High	Medium-high	Low

Remaining Challenges

Cost: better quality means more LLM calls — a tradeoff between API cost and response latency
Explainability: the more complex the pipeline, the harder it is to trace “why this answer came out”
Evaluation criteria: the definition of a “good answer” varies by domain
Domain specialization: there is no universal configuration — tuning is always required

Future Directions

LazyGraphRAG: lightweight graph construction at reduced cost
Multimodal RAG: unified processing of text, images, tables, and charts
Agent collaboration: multiple specialized agents dividing responsibilities
Explainability(XAI): systems that can trace the answer-generation process

Related Categories

🏗 Infrastructure & Architecture — choosing vector DBs and LLM API providers
🛡 AI Governance — hallucination monitoring, answer trustworthiness management

RAG Pipeline Agent Interface