GraphRAGGraphRAG
GraphRAG
从文档里抽实体和关系,建成图,再用图结构回答「跨文档、全局性」的问题。Extract entities and relations into a graph so the system can answer cross-document and global questions.
索引阶段用模型抽取实体、关系与社区摘要,形成知识图谱。
查询阶段先定位相关实体/社区,再沿边扩展上下文。
它补的是向量 RAG 不擅长的「多跳关系」和「整个语料在讲什么」。
Indexing extracts entities, relations and community summaries into a knowledge graph.
Querying locates relevant entities/communities and expands along edges.
It covers multi-hop relations and corpus-level themes that vector RAG misses.
当问题不是找某一段原文,而是「这些文档之间谁和谁有关」时,纯向量检索会漏。When the question is “how are these documents related” rather than “find this passage”, pure vector search falls short.
去导航:找现成的工具、平台、评测与学习资源Directory: ready-made tools, platforms, benchmarks and courses ↗
回到名词地图看第3站的其他 66 个词 →Back to the map: the other 66 terms in station 3 →
- GraphRAG 论文(arXiv)From Local to Global: A Graph RAG Approach to Query-Focused Summarization (arXiv 2404.16130) ↗