====== Microsoft GraphRAG ======
**GraphRAG** is a retrieval-augmented generation approach developed by [[microsoft|Microsoft]] Research that integrates [[knowledge_graphs|knowledge graphs]] with hierarchical community structures to enable AI systems to reason over complex, interconnected data. Introduced in early 2024, GraphRAG substantially outperforms naive vector-based RAG on multi-hop queries, achieving 3.4x accuracy gains (80% vs. 50% correct answers) and 72-83% comprehensiveness scores.(([[https://arxiv.org/abs/2404.16130|Edge, D. et al. "From Local to Global: A Graph RAG Approach to Query-Focused Summarization." arXiv:2404.16130, 2024]]))

  * **Website:** [[https://[[microsoft|microsoft]].[[github|github]].io/graphrag/|microsoft.[[github|github]].io/graphrag]]
  * **[[github|GitHub]]:** [[https://github.com/[[microsoft|microsoft]]/graphrag|github.com/[[microsoft|microsoft]]/graphrag]]
  * **Paper:** [[https://arxiv.org/abs/2404.16130|arXiv:2404.16130]] (2024), Microsoft Research
  * **Version:** 1.0 (December 2024)

===== How It Works =====
GraphRAG extends standard [[retrieval_augmented_generation|RAG]] by building a knowledge graph from source text and layering it with community hierarchies:

=== Indexing (Graph Creation) ===
  - **Text Chunking:** Divide input into TextUnits
  - **Entity/Relationship Extraction:** Use LLMs with domain-specific prompts to extract entities (nodes) and relationships (edges with weights)
  - **Community Detection:** Cluster entities hierarchically using the **Leiden algorithm**
  - **Community Summarization:** Generate LLM summaries for each community covering entities, themes, events, and evidence
  - **Embedding Generation:** Create [[embeddings|embeddings]] for entities, summaries, and TextUnits; store in vector databases (LanceDB, Azure AI Search)

Version 1.0 reduced storage requirements by 43% through an optimized data model.

=== Query Modes ===
^ Mode ^ Description ^ Best For ^
| **Local Search** | Fans out from specific entities to neighbors, retrieves subgraph context | Entity-specific questions ("What about CompanyX?") |
| **Global Search** | Aggregates community summaries for corpus-wide insights | Holistic questions across entire dataset |
| **DRIFT Search** | Hybrid: generates follow-up queries from communities, then iterates via local search | Dynamic multi-hop reasoning |
| **Basic Search** | Standard vector RAG fallback | Simple semantic queries |

===== Python Example =====
<code python>
import asyncio
from pathlib import Path
from graphrag.config import create_graphrag_config
from graphrag.index import create_pipeline_config, run_pipeline
from graphrag.query.indexer_adapters import (
    read_indexer_communities,
    read_indexer_entities,
    read_indexer_reports,
)
from graphrag.query.llm.oai.chat_openai import ChatOpenAI
from graphrag.query.structured_search.global_search.search import GlobalSearch
from graphrag.query.structured_search.global_search.community_context import GlobalCommunityContext

# Step 1: Index documents (run once to build the knowledge graph)
# graphrag index --root ./my_project

# Step 2: Query the indexed graph programmatically
output_dir = Path("./my_project/output")

llm = ChatOpenAI(model="gpt-4o", api_key="your-key")

# Load indexed artifacts
community_reports = read_indexer_reports(
    output_dir / "community_reports.parquet"
)
entities = read_indexer_entities(
    output_dir / "entities.parquet"
)

# Build the global search context from community summaries
context_builder = GlobalCommunityContext(
    community_reports=community_reports,
    entities=entities,
    token_encoder=llm.token_encoder,
)

# Run a global search query across the entire corpus
search_engine = GlobalSearch(
    llm=llm,
    context_builder=context_builder,
    max_data_tokens=12000,
)

async def main():
    result = await search_engine.asearch(
        "What are the main themes and relationships in this dataset?"
    )
    print(result.response)
    print(f"LLM calls: {result.llm_calls}, Tokens: {result.prompt_tokens}")

asyncio.run(main())
</code>

===== Improvements Over Naive RAG =====
Traditional vector-based RAG retrieves individual text chunks by semantic similarity, struggling with:
  * Questions requiring information scattered across multiple documents
  * Multi-hop reasoning chains
  * Global pattern recognition across large corpora

GraphRAG addresses these through:
  * **Graph reasoning:** Connects disparate facts across documents
  * **Hierarchical summaries:** Provide scalable context without token limits
  * **Dynamic community selection:** Automatically picks optimal hierarchy levels (added November 2024)

===== LazyGraphRAG =====
**LazyGraphRAG** (June 2025) optimizes the approach by skipping upfront LLM summarization during indexing. It builds lightweight graphs using NLP techniques and computes summaries dynamically at query time, reducing indexing costs to **0.1% of full GraphRAG** while maintaining query quality.

===== Implementation =====
  * **Pipeline:** Python-based, configurable via YAML/JSON
  * **LLM Support:** GPT-4 Turbo recommended for graph extraction; configurable for other models
  * **Cost Considerations:** Indexing is 100-1000x more expensive than vector RAG (mitigated by LazyGraphRAG)
  * **Auto-tuning:** Domain-specific prompt auto-tuning since September 2024


===== Related Pages =====
  * [[retrieval_augmented_generation|Retrieval-Augmented Generation]]
  * [[llamaindex|LlamaIndex]]
  * [[knowledge_graphs|Knowledge Graphs]]
  * [[tool_augmented_language_models|Tool-Augmented Language Models]]

===== See Also =====
  * [[lightrag|LightRAG]]
  * [[knowledge_graphs|Knowledge Graphs]]
  * [[graph_prompting|Graph Prompting]]
  * [[context_graph|Context Graph]]
  * [[langgraph|LangGraph]]

===== References =====