GraphRAG (Graph Retrieval-Augmented Generation) is an advanced AI framework that combines knowledge graphs with Large Language Models (LLMs) to provide deeper, more context-aware answers than traditional "naive" RAG systems. GraphRAG（图检索增强生成）是一个先进的人工智能框架，将知识图谱与大型语言模型（LLM）结合，提供比传统“天真”RAG系统更深入、更具上下文感知的答案。 

While standard RAG treats information as isolated text chunks and finds them via keyword or vector similarity, GraphRAG models data as a network of nodes (entities like people or places) and edges (relationships between them). 标准RAG将信息视为孤立的文本块，并通过关键词或向量相似度找到它们，而GraphRAG则将数据建模为节点（如人或地点）和边（它们之间的关系）组成的网络。 

Key Capabilities 主要能力

  • Multi-Hop Reasoning: It can answer complex questions that require connecting multiple pieces of information across different documents.
    多跳推理：它可以回答需要将多条信息跨越不同文档连接的复杂问题。
  • Global Summarization: Unlike traditional RAG, which struggles with "What are the main themes of this dataset?", GraphRAG uses hierarchical community detection to summarize entire corpora.
    全局总结：与传统RAG难以理解“该数据集的主要主题是什么？”不同，GraphRAG采用层次式社区检测来总结整个语料库。
  • High Traceability: It provides clear provenance for its answers, allowing users to see exactly which entities and relationships were used to generate a response. 高可追溯性：它为答案提供了明确的来源，使用户能够准确看到哪些实体和关系被用来生成回复。 

Core Workflow 核心工作流程

  1. Indexing: An LLM extracts entities and relationships from raw text to build a Knowledge Graph.
    索引：LLM从原始文本中提取实体和关系，构建知识图谱。
  2. Community Detection: Related entities are grouped into "communities," and summaries are pre-generated for these groups.
    社区检测：相关实体被归入“社区”，并预先生成摘要。
  3. Retrieval: When a query is made, the system traverses the graph to find relevant clusters and paths rather than just looking for similar keywords.
    检索：当查询时，系统会遍历图以寻找相关的聚类和路径，而不仅仅是寻找相似的关键词。
  4. Generation: The LLM synthesizes an answer using the structured relational context. 

Common Use Cases 常见用例

  • Legal & Compliance: Analyzing relationships across thousands of contracts or regulations.
    法律与合规：分析数千份合同或法规中的关系。
  • Biotech & Healthcare: Linking symptoms, drug interactions, and research papers.
    生物技术与医疗保健：症状、药物相互作用与研究论文的关联。
  • Financial Services: Detecting fraud by uncovering suspicious connections or analyzing market trends.
    金融服务：通过发现可疑关联或分析市场趋势来侦测欺诈。
  • Enterprise Search: Solving the "fragmented data" problem where information is spread across siloed documents. 企业搜索：解决信息分散在孤立文档中的“碎片化数据”问题。 

Popular Frameworks 流行框架

  • Microsoft GraphRAG: An open-source pipeline often used for global sense-making across massive datasets.
    Microsoft GraphRAG：一个开源流水线，常用于跨海量数据集进行全球意义构建。
  • Neo4j GraphRAG: A Python package for building applications that leverage the Neo4j graph database.
    Neo4j GraphRAG：一款用于构建利用 Neo4j 图数据库应用的 Python 软件包。
  • Memgraph: An in-memory graph database designed for real-time performance in GraphRAG applications. Memgraph：一种为GraphRAG应用中实时性能设计的内存图数据库。

本文归档于个人知识库 · 领域：CS & AI · 模块：04 Project_CS_AI/LLM