GNN-RAG: Bridging Graph Reasoning and Language Understanding for Knowledge Retrieval
summary
GNN-RAG combines graph neural networks for candidate node classification and path scoring with an LLM for final generation, verbalizing shortest paths from the graph and sending them as prompts; it runs a single-pass GNN reasoning over the KG on a single GPU, reduces LLM calls, and fine-tunes a Llama 2 chat model to outperform GPT‑4 on these benchmarks. Reported gains include 9% F1 uplift on multihop QA and ~15.5% multi-entity accuracy, but it overfits single‑hop queries and struggles with sparse or disconnected graphs.
AI-written summary · about
chapters (11)
- 0:00Introduction to GNN-RAG
- 2:59Understanding LLMs and RAG
- 5:53Knowledge graphs in RAG
- 7:58Graph neural networks explained
- 9:23Multihop question challenges
- 11:49Limitations of traditional RAG
- 13:57How GNN-RAG works
- 16:52GNN-RAG case study
- 18:39Performance improvements with GNN-RAG
- 20:30Limitations and future work
- 21:56Industrial impact and Q&A
related talks
53mPractical GraphRAG - Making LLMs smarter with Knowledge Graphs
20mHybridRAG: A Fusion of Graph and Vector Retrieval
20mPractical GraphRAG: Making LLMs smarter with Knowledge Graphs
26mPractical GraphRAG Making LLMs smarter with Knowledge Graphs
49mPower up LLMs to be Factual Ninjas using GraphRAG
28mEnhancing LLMs with Graph Technology
27mLies, Damn Lies, and AIs
36mAgentic GraphRAG: AI’s Logical Edge