GraphRAG - Knowledge Graph Retrieval-Augmented Generation Cheat Sheet

Updated 2026-05-18

Next Topic: Hyperparameter Tuning Cheat Sheet

🧠Study flashcards on this topic115 cards · spaced repetition→

GraphRAG is an advanced retrieval-augmented generation paradigm that combines knowledge graphs with large language models to address the limitations of standard vector-based RAG. Unlike traditional RAG, which retrieves text chunks via semantic similarity, GraphRAG extracts a structured knowledge graph from documents—capturing entities, relationships, and communities—then uses graph traversal and community summaries to power both local (entity-focused) and global (dataset-wide) reasoning. This enables multi-hop inference, explainable provenance, and improved accuracy on complex queries where answers live in connections, not content. Key to GraphRAG's value is its two-stage architecture: an indexing pipeline that constructs the graph (entity extraction → relationship detection → community clustering → summary generation), and a retrieval pipeline that traverses or queries the graph at inference time. Trade-offs include higher indexing costs (10–100x token usage vs. vanilla RAG) and increased latency, but where relational reasoning matters—finance, healthcare, legal compliance—GraphRAG consistently outperforms embedding-only approaches by 35–46% on multi-hop benchmarks.

What This Cheat Sheet Covers

This topic spans 25 focused tables and 178 indexed concepts, 115 flashcards. Below is a complete table-by-table outline of this topic, spanning foundational concepts through advanced details.

Table 1: Core GraphRAG ConceptsTable 2: GraphRAG vs. Standard RAGTable 3: Entity and Relationship Extraction TechniquesTable 4: Community Detection AlgorithmsTable 5: Query Types and Retrieval PatternsTable 6: Knowledge Graph Construction from DocumentsTable 7: Graph Database OptionsTable 8: Hybrid Vector-Graph ArchitecturesTable 9: Microsoft GraphRAG ImplementationTable 10: LlamaIndex GraphRAG ImplementationTable 11: LightRAG and Alternative ImplementationsTable 12: Text Chunking Strategies for GraphRAGTable 13: Prompting Techniques for GraphRAGTable 14: Query Routing and ClassificationTable 15: Cypher Query Generation and Text2CypherTable 16: Cost Optimization StrategiesTable 17: Performance and ScalabilityTable 18: Monitoring and ObservabilityTable 19: Evaluation Metrics and BenchmarksTable 20: Use Cases and When to Use GraphRAGTable 21: Limitations and ChallengesTable 22: LLM Provider Choices for GraphRAGTable 23: Integration with LangChain and LlamaIndexTable 24: Document Parsing and PreprocessingTable 25: Agentic GraphRAG and Future Directions

Quick IndexSubscribe to unlock

A jump-to index of every table row in this cheat sheet.

Mind MapSubscribe to unlock

An interactive map of every table and concept in this topic.

Table 1: Core GraphRAG Concepts

GraphRAG fundamentally reimagines retrieval by replacing flat semantic search with structured graph reasoning. Understanding these foundational concepts clarifies why GraphRAG excels at relationship-driven queries where standard RAG fails.

Concept	Example	Description
GraphRAG	Microsoft's approach: extract entities → build hierarchy → generate summaries → query via map-reduce	• RAG paradigm that uses knowledge graphs instead of vector embeddings for retrieval • enables multi-hop reasoning and explainable answers
Knowledge Graph (KG)	Nodes = `Person`, `Organization`; Edges = `WORKS_FOR`, `INVESTED_IN`	• Structured representation of data as entities (nodes) connected by relationships (edges) • captures semantics beyond flat text
Entity Extraction	LLM extracts `"John Smith, CEO, TechCorp"` → nodes `Person(John Smith)`, `Organization(TechCorp)`, edge `ROLE_AT`	• Process of identifying named entities (people, places, concepts) from unstructured text • forms graph nodes
Relationship Extraction	From text: `"Alice hired Bob"` → triple `(Alice, HIRED, Bob)`	• Detecting semantic connections between entities • forms graph edges • can be LLM-based or NLP rule-based
Community Detection	Leiden algorithm clusters related entities into communities	• Graph clustering to group densely connected entities • enables hierarchical summarization at scale
Community Summary	LLM generates: `"Community 7 focuses on AI safety research, key members: Anthropic, OpenAI..."`	• Abstract of a detected community • generated by LLM from entities/relationships • powers global search
Local Search	Query: `"What did Alice say about X?"` → retrieves Alice's neighborhood + related entities	Retrieval pattern that starts from specific entities found via embedding similarity, then expands via graph traversal
Global Search	Query: `"What are the main themes?"` → map-reduce over all community summaries	• Retrieval pattern for holistic questions • aggregates insights from community reports via map-reduce
DRIFT Search	Combines global search (broad answer) + local search (entity-specific follow-ups) in multi-stage flow	Hybrid retrieval that dynamically switches between global and local patterns based on query
Indexing Pipeline	Text chunking → entity extraction → graph construction → community detection → summary generation	Offline phase that transforms raw documents into a queryable knowledge graph structure
Retrieval Pipeline	User query → embedding → entity matching → graph traversal → context assembly → LLM generation	Online phase that retrieves relevant graph context and generates answers at inference time
Graph Traversal	Starting from `Person(Alice)`, traverse `KNOWS` edges 2 hops → retrieve friends-of-friends	• Following edges to expand context • enables multi-hop reasoning unreachable by vector similarity alone

Table 2: GraphRAG vs. Standard RAG

Standard RAG and GraphRAG share the goal of grounding LLMs in external knowledge, but their architectures differ fundamentally. This table highlights where each approach excels and where trade-offs emerge.

Dimension	Example	Description
Vector RAG	Chunks → embeddings → vector DB → top-K similarity search for retrieval	• Uses semantic similarity over text embeddings • fast, simple, works well for single-hop factual recall

The rest of this section is part of the full cheat sheet

Table 3: Entity and Relationship Extraction Techniques

The quality of a GraphRAG system hinges on how accurately entities and relationships are extracted from text. These methods range from rule-based NLP to fully LLM-driven approaches.

Method	Example	Description
LLM-Based Extraction	Prompt: `"Extract entities and relationships from: [text]"` → structured JSON output	• Uses prompted LLMs (GPT-4, Claude) to identify entities and edges • flexible but token-intensive

The rest of this section is part of the full cheat sheet

Table 4: Community Detection Algorithms

Graph-RAG relies on community detection to hierarchically cluster entities, enabling scalable summarization. These algorithms partition the graph into densely connected groups.

Algorithm	Example	Description
Leiden Algorithm	Refines Louvain by ensuring well-connected communities; avoids disconnected subgraphs	• Modularity-based clustering • gold standard in GraphRAG for community hierarchy • faster and more reliable than Louvain

The rest of this section is part of the full cheat sheet

Table 1: Core GraphRAG Concepts

Concept	Example	Description
GraphRAG	Microsoft's approach: extract entities → build hierarchy → generate summaries → query via map-reduce	• RAG paradigm that uses knowledge graphs instead of vector embeddings for retrieval • enables multi-hop reasoning and explainable answers
Knowledge Graph (KG)	Nodes = `Person`, `Organization`; Edges = `WORKS_FOR`, `INVESTED_IN`	• Structured representation of data as entities (nodes) connected by relationships (edges) • captures semantics beyond flat text
Entity Extraction	LLM extracts `"John Smith, CEO, TechCorp"` → nodes `Person(John Smith)`, `Organization(TechCorp)`, edge `ROLE_AT`	• Process of identifying named entities (people, places, concepts) from unstructured text • forms graph nodes
Relationship Extraction	From text: `"Alice hired Bob"` → triple `(Alice, HIRED, Bob)`	• Detecting semantic connections between entities • forms graph edges • can be LLM-based or NLP rule-based
Community Detection	Leiden algorithm clusters related entities into communities	• Graph clustering to group densely connected entities • enables hierarchical summarization at scale
Community Summary	LLM generates: `"Community 7 focuses on AI safety research, key members: Anthropic, OpenAI..."`	• Abstract of a detected community • generated by LLM from entities/relationships • powers global search
Local Search	Query: `"What did Alice say about X?"` → retrieves Alice's neighborhood + related entities	Retrieval pattern that starts from specific entities found via embedding similarity, then expands via graph traversal
Global Search	Query: `"What are the main themes?"` → map-reduce over all community summaries	• Retrieval pattern for holistic questions • aggregates insights from community reports via map-reduce
DRIFT Search	Combines global search (broad answer) + local search (entity-specific follow-ups) in multi-stage flow	Hybrid retrieval that dynamically switches between global and local patterns based on query
Indexing Pipeline	Text chunking → entity extraction → graph construction → community detection → summary generation	Offline phase that transforms raw documents into a queryable knowledge graph structure
Retrieval Pipeline	User query → embedding → entity matching → graph traversal → context assembly → LLM generation	Online phase that retrieves relevant graph context and generates answers at inference time
Graph Traversal	Starting from `Person(Alice)`, traverse `KNOWS` edges 2 hops → retrieve friends-of-friends	• Following edges to expand context • enables multi-hop reasoning unreachable by vector similarity alone

Table 2: GraphRAG vs. Standard RAG

Dimension	Example	Description
Vector RAG	Chunks → embeddings → vector DB → top-K similarity search for retrieval	• Uses semantic similarity over text embeddings • fast, simple, works well for single-hop factual recall

Table 3: Entity and Relationship Extraction Techniques

The quality of a GraphRAG system hinges on how accurately entities and relationships are extracted from text. These methods range from rule-based NLP to fully LLM-driven approaches.

Method	Example	Description
LLM-Based Extraction	Prompt: `"Extract entities and relationships from: [text]"` → structured JSON output	• Uses prompted LLMs (GPT-4, Claude) to identify entities and edges • flexible but token-intensive

Table 4: Community Detection Algorithms

Graph-RAG relies on community detection to hierarchically cluster entities, enabling scalable summarization. These algorithms partition the graph into densely connected groups.

Algorithm	Example	Description
Leiden Algorithm	Refines Louvain by ensuring well-connected communities; avoids disconnected subgraphs	• Modularity-based clustering • gold standard in GraphRAG for community hierarchy • faster and more reliable than Louvain

GraphRAG – Knowledge Graph Retrieval-Augmented Generation Cheat Sheet

Table 1: Core GraphRAG Concepts

Table 2: GraphRAG vs. Standard RAG

Table 3: Entity and Relationship Extraction Techniques

Table 4: Community Detection Algorithms

GraphRAG – Knowledge Graph Retrieval-Augmented Generation Cheat Sheet

Table 1: Core GraphRAG Concepts

Table 2: GraphRAG vs. Standard RAG

Table 3: Entity and Relationship Extraction Techniques

Table 4: Community Detection Algorithms