When Vectors Break Down: Graph-Based RAG for Dense Enterprise Knowledge
summary
Claims 86.31% accuracy on RobustQA with sub‑second responses and outperforms seven vector systems; uses a small specialized LLM to map documents into nodes/edges, context‑aware chunking and compression, stores graph JSON in a Lucene search engine for scale, and applies fusion‑in‑decoder over retrieved passages to lower hallucination and enable multihop reasoning. Benchmarks and a published white paper back the setup and show practical gains on product, finance, and technical doc collections.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 0:28Vector Search Limitations
- 1:43Graph-based RAG Overview
- 3:42Enterprise Data Challenges
- 5:08From Vector to Graph RAG
- 7:38Technical Challenges
- 9:30KG-FID Architecture
- 13:06Benchmarking Results
- 13:45Conclusion
related talks
27mWhen Vectors Aren’t Enough: Unlocking AI’s Hidden Connections with Graphs
30mBuilding a HybridRAG Document Question-Answering System
53mPractical GraphRAG - Making LLMs smarter with Knowledge Graphs
28mEnhancing LLMs with Graph Technology
23mMeasuring the Performance of VectorRAG vs GraphRAG
27mLies, Damn Lies, and AIs
36mAgentic GraphRAG: AI’s Logical Edge
30mAgentic RAG with Knowledge Graphs Explained