Zilliz open-sources vector-graph-rag for hybrid retrieval
Zilliz has open-sourced vector-graph-rag, a project that adds knowledge-graph recall to conventional vector RAG and combines vector similarity with entity relationships. The Juejin article describes a five-layer architecture using Milvus for vector indexing, plus entity normalization, configurable multi-hop retrieval, evidence validation, and source-linked answers. The author reports improved complex-question accuracy, but also warns that extraction quality, ongoing graph maintenance, added system complexity, and integration work remain important considerations.
The project gives enterprise RAG teams an incremental way to add entity disambiguation, relationship-based retrieval, and evidence chains to existing Milvus-based systems, particularly for multi-hop and relation-heavy questions.
最近一直在做企业 AI 知识库私有化落地，也持续在账号中给大家分享高价值的 AI 开源项目。
做AI知识库私有化交付这两年，我最深的体感是：纯向量RAG的天花板，已经越来越明显了。
早期项目里，客户需求很简单：把文档传上去，能关键词搜索、能简单问答，就足够满意。但随着项目深入，需求很快就往复杂走：问 "张三负责的项目有哪些合作供应商"" 去年 Q3 出现过的同类型设备故障有哪些解决方案 ""这份合同和哪份制度条款存在冲突"—— 这类需要跨文档、找关联、做多跳推理的问题，纯向量 RAG 几乎次次翻车。
不是切片拆得不够细，也不是 Embedding 模型不够好，而是底层逻辑就有局限：纯向量检索本质是 "片段相似度匹配"，它只认文字像不像，不理解实体和实体之间的关系。就像一个人只看过一堆零散纸条，知道每句话写了什么，但分不清谁是谁、事情之间有什么联系，遇到需要串起来的问题，只能靠猜，幻觉自然就来了。