GraphRAG

1. Project/Product Overview

2. What does it mostly do?

The core mission of the GraphRAG is to solve the essential limitation of the traditional vector RAG *: vector retrieval can only return text fragments that are semantically similar to the question, but cannot understand the relationship between the fragments, let alone answer the macro question that "you need to look at all the data to answer.

Core Innovation: Knowledge Graph RAG

Flow of traditional vector RAG:

文档 → 文本分块 → 向量嵌入 → 余弦相似度检索 → 丢给 LLM 生成

GraphRAG process:

文档 → TextUnit 切分 → LLM 实体/关系抽取 → 知识图谱构建 → Leiden 社区检测 → 层次化社区摘要 → LLM 生成

The essential difference between the two lies in the intermediate representation: vector RAG turns the document into "a bunch of numbers" and GraphRAG turns the document into "a structured network of relationships". The former can only answer "which text is most similar to your question", while the latter can answer "what the entire data set says and how entities are related to each other".

Four search modes (v3.x full version)

Index Pipeline Details

The GraphRAG indexing process is the core and most expensive part:

1. TextUnit segmentation: The original document is divided into analyzable units (TextUnit) as the granularity unit for subsequent processing, and fine-grained references in the output are provided at the same time.
2. Entity/Relationship Extraction:LLM automatically identifies entities (people, organizations, places, concepts, etc.) and their relationships from each TextUnit. This is the most costly part-each piece of text requires multiple LLM calls
3. Knowledge Graph Construction: Entities are used as nodes and relationships are used as edges with types to build a global knowledge graph.
4. Leiden Community Detection : Hierarchical clustering of the map-each circle represents an entity, the size represents the degree, and the color represents the community to which it belongs. Generate a multi-level community structure
5. Community Summary Generation: Generate a natural language summary for each community from the bottom up. The bottom community describes the details, and the upper community describes the broader theme.

文本输入 → [实体抽取] → 实体/关系 → [图构建] → 知识图谱
                                                    ↓
查询输出 ← [LLM 生成] ← 检索上下文 ← [查询引擎] ← 社区摘要 ← [层次化摘要] ← 社区结构

Prompt Tuning

Microsoft strongly recommends Prompt Tuning when using GraphRAG-the system provides an automated prompt tuning mechanism to adjust the prompt of entity extraction, community summary and other links according to your data characteristics. The out-of-the-box default prompt may not work well in specific areas.

3. Applicable Scenario

4. Not quite the scene

5. Core Competence List

5.1 Index Layer Capability

5.2 query layer capability

5.3 Integration and Ecology

6. Architecture/deployment/integration approach

Installation Method

# 从 PyPI 安装
python -m pip install graphrag

# 初始化项目
graphrag init

Deployment Mode

LLM Integration

-Direct support:OpenAI API(GPT-4o, GPT-4.1, etc.), Azure OpenAI (including Managed Identity authentication)

-Configurable access: any service compatible with OpenAI API format (such as Ollama local model, DeepSeek, Tongyi Qiwen, etc.)

-Embed Model:OpenAI text-embedding-3-small/large or Azure OpenAI Embed

-Key Tip: Different LLMs have a significant impact on the quality of entity extraction, and Microsoft recommends using GPT-4-level models to ensure the quality of the map.

Configuration file (settings.yaml)

After installation, 'settings.yaml' is automatically generated to control all indexes and query parameters:

-LLM model selection and parameters (chat model embedding model)

-Text block size and overlap

-prompt template for entity/relationship extraction

-Leiden parameters for community detection

-Output format (Parquet file)

How to use #7.

Complete Example: From Installation to Query

Step 1: Environment Preparation and Installation

# 创建项目空间
mkdir graphrag_demo && cd graphrag_demo
python -m venv .venv
source .venv/bin/activate  # Windows: .venv\Scripts\activate

# 安装 GraphRAG
python -m pip install graphrag

Step 2: Initialize project

graphrag init

will be generated in the current directory:

-'input/'-put the text file to be indexed

-'.env'-Fill in 'GRAPHRAG_API_KEY = '

-'settings.yaml'-configuration of indexes and queries

-'prompts/'-prompt template for each link

Step 3: Prepare data and index

# 下载示例文本（狄更斯《圣诞颂歌》）
curl https://www.gutenberg.org/cache/epub/24022/pg24022.txt -o ./input/book.txt

# 执行索引（注意：会消耗 LLM API 额度！）
graphrag index

After the index is completed, a series of Parquet files are generated in the 'output/'directory, including entity tables, relationship tables, community tables, community report tables, text unit tables, and so on.

Step 4: Query

# 全局查询 — 询问整个数据集的主题
graphrag query "What are the top themes in this story?"

# 局部查询 — 询问特定实体的详细信息
graphrag query "Who is Scrooge and what are his main relationships?" --method local

# DRIFT 搜索 — 结合局部和全局上下文
graphrag query "What is the significance of the Ghost of Christmas Past?" --method drift

# 基础搜索 — 传统向量检索模式
graphrag query "What year was this story published?" --method basic

Python API example:

import asyncio
from graphrag.api import build_index
from graphrag.query.context import set_context
from graphrag.query.engine import QueryEngine

async def main():
    # 执行索引（从配置文件读取参数）
    await build_index(root_dir="./graphrag_demo")

    # 设置查询上下文并加载索引
    await set_context(root_dir="./graphrag_demo")

    # 创建查询引擎
    engine = QueryEngine()

    # 全局搜索
    result = await engine.global_search("这个数据集涵盖了哪些主要主题？")
    print(result)

    # 局部搜索
    result = await engine.local_search("主角与哪些角色有重要关系？")
    print(result)

asyncio.run(main())

Cost Warning: The above example uses about 30000 words of "A Christmas Carol" and costs about $0.3-0.5 to index in GPT-4o. If a 32000-word book is indexed with GPT-4 Turbo, the community actually costs about $6 to $7. It is recommended to start with a minimal data set to verify the effect.

8. What can I say before sales

8.1 a sentence positioning

• * "GraphRAG redefines RAG with knowledge graph-so that your AI can not only 'fath' relevant information, but also 'understand' the global structure and correlation of information. "**

8.2 customer pain points → solutions

8.3 Differentiated Selling Points

vs traditional vector RAG(Haystack, LlamaIndex, etc.):

-Vector RAG is the match of "like or not", GraphRAG is the understanding of "what relationship"-the quality level of the answer is different

-Vector RAG is suitable for factual questions and answers ("what is stipulated in article number of the articles of association") and GraphRAG for analytical questions and answers ("what is the cooperative relationship between various departments of the company")

-Vector RAG index cost is extremely low (only one embedding is required),GraphRAG index cost is 10 to 100 times higher, but the answer quality is an order of magnitude higher on macro questions

-GraphRAG Basic Search mode is to return to vector RAG, which can be mixed

vs LightRAG (HKU, October 2024):

-LightRAG abandoned community detection and hierarchical summarization (GraphRAG the most expensive part) and used graph embedding low-level/high-level dual retrieval instead

-The LightRAG index cost is about 1/100 of the GraphRAG, the query speed is much faster, and it supports real incremental updates

-But the LightRAG lacks global community-level understanding-it can only find related entities and relationships and cannot generate a "thematic overview of the entire corpus"

-GraphRAG Global Search is truly "overlooking the big picture" and LightRAG can't do it

-Evaluation and comparison: GraphRAG have obvious advantages in Contextual Summarize tasks (64.40 vs LightRAG 48.85, from GraphRAG-Bench data)

-Pre-sales conclusion: If you need macro thematic analysis, choose GraphRAG; if you need low-cost, high-frequency updated entity-level Q & A, choose LightRAG

vs LazyGraphRAG (Microsoft itself, November 2024):

-LazyGraphRAG is Microsoft's own lightweight variant, with indexing costs reduced to 0.1 percent of standard GraphRAG

-Higher query latency at the cost of deferring a lot of LLM work to queries

-Suitable for scenarios where "global understanding is occasionally required, but daily queries are mainly local"

-This is the most acceptable entry point for customers: first verify the value with LazyGraphRAG, and then upgrade to the full version according to requirements.

vs Pure Knowledge Graph Scheme (Neo4j Cypher):

-GraphRAG automatic construction of graphs from unstructured text without manual ontology design (Ontology)

-Traditional knowledge mapping requires domain experts to manually define schemas, which is costly and less flexible.

-The GraphRAG map is generated by LLM, the quality depends on the LLM ability, there may be extraction errors.

-For customers who already have structured data, traditional graph database queries are more accurate and have lower latency

8.4 Customer Value Story Line

1. Cut in:"Can your current RAG system answer the question 'what is the dependency relationship between all the projects in the company'? No way, right? Because traditional RAG is only good at 'finding similar pieces' and not good at 'understanding the global relationship '."

1. Resonance :"Your analysts have to read hundreds of reports every day to write summaries. RAG system can only help you find paragraphs. Finally, the human brain is doing the work of connecting various information points. It's not at all what AI should do."

1. Demo : Take out a real customer document (such as annual report set) and use GraphRAG to do Global Search → Automatically generate topic overview, entity relationship network diagram and key entity analysis. Let customers see firsthand the power of "automatically extracting this diagram from hundreds of documents.

1. Advanced : Starting from the LazyGraphRAG of a low-cost entry point → Gradually upgrade to the full version after verifying the value → Four search modes are used on demand (Local/Basic for daily use and Global/DRIFT for macro analysis) → Incremental update to ensure data freshness.

1. Heavy *:"Microsoft Research, 34,000 GitHub Stars,MIT Open Source Protocol. The paper was published at EMNLP and other top meetings, and the community was active and updated frequently. This is not an experimental project, but a production-level tool that has been relied upon by 500 projects."

9. Frequently Asked Customer Questions

10. PoC Recommendations

Recommended PoC Direction: Macro Topic Analysis and Intelligent Question Answering of Enterprise Document Set

Total: Approximately 4.5 business days (excluding LLM API approval time)

Validation Metrics:

-Global Search's answer is comprehensive: manual evaluation> 4/5 (traditional RAG is usually only 2~3/5)

-Local Search entity relationship accuracy:> 80%

-entity extraction accuracy: spot check 50 entities, accuracy> 85%

-Index cost: record the actual token consumption and make budget reference for production deployment.

Cost Estimate (PoC Phase):

-200 documents (about 500000~1 million Chinese characters) are indexed in GPT-4o-mini: about $1~5

-Prompt Tuning and repeated testing: about $3~10

-Total PoC LLM cost: about $5~15 (very controllable)

11. Risks and Considerations

12. My Pre-Sales Judgment

Recommendation: Highly recommended for specific scenarios (suitable for large-scale unstructured document scenarios that require "global understanding of entity relationship analysis")

Reason:

1. Methodology Leading:GraphRAG is not an incremental improvement of vector RAG, but a paradigm transition from "fragment matching" to "global understanding. In business scenarios that require a panoramic view, it provides value that traditional RAGs cannot match.
2. Microsoft Brand + Academic Authority : Microsoft Research Institute, Top Paper, 34K Stars -- Trust Endorsement is strong enough to convince technical decision makers.
3. MIT open source: no commercial lock-in risk, customers can freely use, modify, internal distribution.
4. Four search modes are flexible: from zero-cost Basic Search to full-featured Global Search, choose and mix on demand.
5. LazyGraphRAG lowers barriers to entry : Cost-sensitive customers can start with a light version of 0.1 per cent cost and verify the value before upgrading.
6. Ecology Gradually Mature :FalkorDB and other manufacturers have launched production-level SDK,GraphRAG-Bench providing standardized evaluation and active community.

Recommended Customer Persona:

-Research/Intelligence Agencies: Need to discover topics, trends and associations from large volumes of unstructured text

-Financial/Legal/Compliance: Cross-document entity relationship analysis, due diligence, compliance review

-Large enterprise knowledge management: global understanding of internal documents, cross-departmental knowledge association

-Have an LLM API budget (hundreds to thousands of dollars per month LLM fees are acceptable)

-Technical team capable of Python-level integration and tuning

Not recommended situations:

-Core requirements are just simple Q & A/FAQ (vector RAG is more efficient and cheaper)

-Extremely limited budget or allergic to the cost of LLM API calls

-Real-time dialog systems that require extremely low latency (Global Search latency is too high)

-The amount of data is very small (<100 documents, the map advantage is not obvious)

-Team lacks Python development capabilities (requires code-level integration and tuning)

-Official SLA/commercial support required (GraphRAG research projects, non-commercial products)

13. REFERENCE

-GitHub repository: https://github.com/microsoft/graphrag

-Official Document: https://microsoft.github.io/graphrag

-Academic paper: https://arxiv.org/pdf/2404.16130

-Microsoft Research Blog: https://www.microsoft.com/en-us/research/blog/graphrag-unlocking-llm-discovery-on-narrative-private-data/

-PyPI:https://pypi.org/project/graphrag/

-LazyGraphRAG papers: https://arxiv.org/abs/2411.14743

-LightRAG (competition comparison):https://github.com/HKUDS/LightRAG

-FalkorDB GraphRAG SDK:https://github.com/FalkorDB/graphrag-sdk

-GraphRAG-Bench evaluation: https://graphrag-bench.github.io/

-GraphRAG Cost Analysis Reference: https://aiwiki.ai/wiki/graphrag

• analysis date: 2026-07-02 | data aging: GitHub information is pulled in real time, product functions are based on official document v3.1.0 *