停止使用简单的RAG——为AI上下文添加关系
我最近一直在大量研究RAG系统，并不断遇到同样的问题：它们能检索到相关的片段，但却失去了它们之间的关系。<p>当处理真实系统（如文档、API、基础设施）时，这很快就会成为一个问题，因为理解事物之间的连接比仅仅找到相似文本更为重要。<p>因此，我构建了Mindex：
https://usemindex.dev<p>它将语义搜索与知识图谱层结合在一起，因此它不仅返回孤立的片段，还能连接文档并展示它们之间的关系。<p>它通过命令行界面（CLI）和MCP工作，因此您可以直接将其集成到像Claude Code、Cursor或您自己的代理工具中。<p>我还添加了一个视觉比较，展示了简单RAG与基于图谱的检索之间的区别，使差异更加明显。<p>非常希望能得到反馈，尤其是来自那些使用RAG或从事开发工具工作的人。
查看原文
I’ve been working a lot with RAG systems recently, and kept running into the same issue: they retrieve relevant chunks, but lose the relationships between them.<p>This becomes a problem pretty quickly when dealing with real systems (docs, APIs, infra), where understanding how things connect matters more than just finding similar text.<p>So I built Mindex:
https://usemindex.dev/<p>It combines semantic search with a knowledge graph layer, so instead of returning isolated chunks, it can connect documents and surface how they relate.<p>It works via CLI and MCP, so you can plug it directly into tools like Claude Code, Cursor, or your own agents.<p>I also added a visual comparison showing naive RAG vs graph-based retrieval, which makes the difference clearer.<p>Would love feedback — especially from people building with RAG or working on developer tooling.