GraphStuff.FM: The Neo4j Graph Database Developer Podcast Getting the Word out on Knowledge Graphs with Leann Chen
Jun 1, 2024
Leann Chen from Diffbot discusses extracting web data for a knowledge graph. They explore integrating knowledge graph builders with Neo4j, discuss vectors, and DSPY framework. The episode also covers using language models and knowledge graphs in content creation, interactive graph visualization tools, Langchain, and NODES 2024 proposals call.
AI Snips
Chapters
Transcript
Episode notes
Hybrid Vector+Graph RAG Is Powerful
- Combining graph grounding and vector embeddings can give RAG systems both determinism and semantic richness.
- Neo4j supports embedding storage on nodes, enabling hybrid vector+graph RAG workflows.
Use DSPy But Validate Its Outputs
- Try modular frameworks like DSPy to reduce manual prompt engineering and enable self-improving prompts.
- Expect unpredictability and validate DSPy results per task rather than assuming universal improvements.
When Self-Improving Prompts Go Off-Track
- Leann shared a DSPy example where an automatic prompt improvement changed "SpaceX" into a generic "companies with Elon Musk," losing specificity.
- That showed self-improvement can drift away from the original intent and require oversight.