---
title: "𝗬𝗼𝘂 𝗰𝗮𝗻 𝗱𝗼 𝗴𝗿𝗮𝗽𝗵 𝘁𝗿𝗮𝘃𝗲𝗿𝘀𝗮𝗹 𝘄𝗶𝘁𝗵𝗼𝘂𝘁 𝗮 𝗴𝗿𝗮𝗽𝗵 𝗱𝗮𝘁𝗮𝗯𝗮𝘀𝗲. Here..."
source_name: "Milvus(@milvusio)"
original_url: "https://x.com/milvusio/status/2049874919710720152"
canonical_url: "https://www.traeai.com/articles/6cb4d00a-322c-4339-80f2-ce53848ab531"
content_type: "tweet"
language: "中文"
score: 7.5
tags: ["Milvus","知识图谱","图遍历","向量数据库","机器学习"]
published_at: "2026-04-30T15:34:00+00:00"
created_at: "2026-05-01T02:16:50.900282+00:00"
---

# 𝗬𝗼𝘂 𝗰𝗮𝗻 𝗱𝗼 𝗴𝗿𝗮𝗽𝗵 𝘁𝗿𝗮𝘃𝗲𝗿𝘀𝗮𝗹 𝘄𝗶𝘁𝗵𝗼𝘂𝘁 𝗮 𝗴𝗿𝗮𝗽𝗵 𝗱𝗮𝘁𝗮𝗯𝗮𝘀𝗲. Here...

Canonical URL: https://www.traeai.com/articles/6cb4d00a-322c-4339-80f2-ce53848ab531
Original source: https://x.com/milvusio/status/2049874919710720152

## Summary

Milvus 提出了一种无需图数据库即可进行图遍历的方法，通过向量图 RAG，将知识图谱三元组嵌入 Milvus 向量数据库中，实现高效查询与子图扩展，适用于多跳推理任务。

## Key Takeaways

- 向量图 RAG 跳过了传统图数据库，直接在 Milvus 中存储和查询实体及关系的向量表示。
- 查询过程涉及从问题中提取实体、向量搜索匹配、子图扩展以及 LLM 重排序，以生成答案。
- 该方法在 MuSiQue、HotpotQA 和 2WikiMultiHopQA 基准测试上平均 Recall@5 达到 87.8%，并提供 MIT 开源许可。

## Content

Title: Milvus on X: "𝗬𝗼𝘂 𝗰𝗮𝗻 𝗱𝗼 𝗴𝗿𝗮𝗽𝗵 𝘁𝗿𝗮𝘃𝗲𝗿𝘀𝗮𝗹 𝘄𝗶𝘁𝗵𝗼𝘂𝘁 𝗮 𝗴𝗿𝗮𝗽𝗵 𝗱𝗮𝘁𝗮𝗯𝗮𝘀𝗲. Here's how.

A knowledge graph is triplets: (Einstein, developed, Relativity). Typically you store them in Neo4j, traverse edges, and reason over connections.

Vector Graph RAG skips https://t.co/6NwBdjrXrt" / X

URL Source: http://x.com/milvusio/status/2049874919710720152

Markdown Content:
## Post

## Conversation

𝗬𝗼𝘂 𝗰𝗮𝗻 𝗱𝗼 𝗴𝗿𝗮𝗽𝗵 𝘁𝗿𝗮𝘃𝗲𝗿𝘀𝗮𝗹 𝘄𝗶𝘁𝗵𝗼𝘂𝘁 𝗮 𝗴𝗿𝗮𝗽𝗵 𝗱𝗮𝘁𝗮𝗯𝗮𝘀𝗲. Here's how. A knowledge graph is triplets: (Einstein, developed, Relativity). Typically you store them in Neo4j, traverse edges, and reason over connections. Vector Graph RAG skips the graph database entirely. It embeds every triplet into Milvus. Entities become vectors. Relations become vectors. The whole knowledge graph lives in one vector database. At query time: • Extract entities from the question. • Vector search finds the closest matching entities and relations. • 𝗦𝘂𝗯𝗴𝗿𝗮𝗽𝗵 𝗲𝘅𝗽𝗮𝗻𝘀𝗶𝗼𝗻: if (Einstein → developed → Relativity) is retrieved, the system fetches neighboring triplets like (Relativity → revolutionized → Physics). • 𝗢𝗻𝗲 𝗟𝗟𝗠 𝗰𝗮𝗹𝗹 𝗿𝗲𝗿𝗮𝗻𝗸𝘀 the collected context. Answer generated. Two LLM calls total per query. 87.8% avg Recall@5 on standard multi-hop benchmarks (MuSiQue, HotpotQA, 2WikiMultiHopQA). pip install vector-graph-rag — runs locally with Milvus Lite. MIT licensed, fully open source. ![Image 1: 🔗](https://abs.twimg.com/emoji/v2/svg/1f517.svg)Try it on GitHub: [github.com/zilliztech/vec](https://t.co/uyBstuG3mS)

![Image 2](https://pbs.twimg.com/amplify_video_thumb/2049074750602067968/img/AsLRULhjn9XUbY3x.jpg)