Enhancing RAG-based apps by constructing and leveraging knowledge graphs with open-source LLMs
summary
Demo builds a Metallica dataset, stores combined text embeddings in LanceDB with gte‑tiny embeddings and shows vector RAG returns correct song answers sometimes but ranks targets second or third. A Kùzu graph stores artist album song nodes and LangChain / Llama 3.1 (405B) prompts generate Cypher queries; one prompt omission produces a wrong query. Combining a local llama‑cpp pipeline, Outlines for structured graph extraction from text, and a simple yes/no classifier yields a hybrid GraphRAG that works reasonably well but still makes occasional errors.
AI-written summary · about
chapters (10)
- 0:00Enhancing RAG with knowledge graphs
- 1:31Vector-only retrieval limits
- 5:21Augmented generation with LLMs
- 8:00Vector-only limits for structured queries
- 9:14Graph-only retrieval with Kùzu
- 12:37Generating Cypher queries with LLMs
- 17:28Combining vector and graph retrieval
- 20:29Building knowledge graphs from text
- 23:00Conclusion and key takeaways
- 24:36Q&A
related talks
53mPractical GraphRAG - Making LLMs smarter with Knowledge Graphs
26mPractical GraphRAG Making LLMs smarter with Knowledge Graphs
20mHybridRAG: A Fusion of Graph and Vector Retrieval
49mPower up LLMs to be Factual Ninjas using GraphRAG
28mEnhancing LLMs with Graph Technology
27mLies, Damn Lies, and AIs
01h 12mBuild RAG using Llama 3.1
20mPractical GraphRAG: Making LLMs smarter with Knowledge Graphs