# 清华团队提出KG-Infused RAG：用知识图谱激活检索

- 来源：OpenBMB (@OpenBMB)
- 发布时间：2026-07-21 21:00
- AIHOT 分数：54
- AIHOT 链接：https://aihot.virxact.com/items/cmruoof1z01gnbiufh8ltuff0
- 原文链接：https://x.com/OpenBMB/status/2079551973024473553

## AI 摘要

清华NLP与东北大学提出KG-Infused RAG，将知识图谱（Wikidata5M，约2100万三元组）引入RAG，通过“扩散激活”机制同时检索文本段落与结构化事实。

## 正文

When you try to recall something， one concept pulls up a whole web of related ones， even things the question never mentioned. Cognitive science calls this spreading activation. Most RAG systems have none of it， and they stay locked to a single source， either text passages or a knowledge graph， never both.🤔

KG-Infused RAG from @TsinghuaNLP （OpenBMB member） and Northeastern University brings knowledge graphs into RAG and uses spreading activation to drive both retrieval and generation， so answers rest on text passages and structured facts at once.

Paper： KG-Infused RAG： Augmenting Corpus-Based RAG With External knowledge Graphs
arxiv： https://arxiv.org/abs/2506.09542
Code： https://github.com/thunlp/KG-Infused-RAG

1️⃣It starts from the query's seed entities and hops outward over a curated KG （Wikidata5M， ~21M triples）. Each round has the LLM pick relevant triples from the current entities' one-hop neighbors， accumulate them into a growing subgraph， and track already-activated entities to avoid loops， covering one-hop to k-hop neighborhoods before summarizing the subgraph into natural language.

2️⃣Those activated facts are used twice. First they expand the query （simplifying a complex question by swapping intermediate entities， or adding clues that text alone matches poorly）， then dual retrieval merges passages from both the original and expanded queries. At generation， retrieved passages are summarized into a query-focused note that the KG subgraph summary then augments into a fact-enhanced note.

3️⃣ It works， and the structured knowledge is the reason. On four multi-hop datasets across Qwen2.5-7B and LLaMA3.1-8B， KG-Infused RAG beats every standard baseline； with two activation rounds and DPO， the DPO-trained LLaMA3.1-8B improves Avg over Vanilla RAG by 3.8% to 13.8%. Plain query expansion without KG facts barely moves， confirming external structure is what matters.

4️⃣It is plug-and-play. Dropped into Self-RAG it adds another 5.1% to 16.9% Avg across four datasets， and viewed purely as retrieval， KG-Based query expansion lifts R@6 by 4.0% to 8.4%. The whole thing uses existing human-built graphs， avoiding both the cost and the extraction noise of building graphs from text on the fly.