Building Knowledge Graph-Based Agents with Structured Text Generation
summary
Builds a knowledge graph from plain text by forcing LLM outputs into a pydantic JSON schema and masking token probabilities with outlines then stores nodes and edges in kutsu and runs Cypher queries; uses Lama CPP for inference and L chain only to extract the DB schema; defines tools query_generator, query_checker and conversational_response and runs a thought/action/observation loop as agent logic; demo extracts person/location/discovery triples (Marie Curie → polonium radium), expert-curation step refines types, and the notebook demo works within seconds.
AI-written summary · about
chapters (10)
- 0:00Welcome
- 1:05Presenter and problem
- 3:53Structured output vs generation
- 6:45Structured generation with regex
- 9:18Forcing correct JSON outlines
- 12:50Building knowledge graphs
- 19:13Designing reasoning agents
- 24:29Agent execution demo
- 24:51Summary
- 25:55Q&A
related talks
32mMulti-Agent Query Generation w/ LLMs on Complex Knowledge Graphs
29mKnowledge graph based chatbots
28mAgentic GraphRAG: Multi-Agent Knowledge Graph
19mWisdom-Driven Knowledge Augmented Generation at Scale
21mAgentic GraphRAG: Autonomous Knowledge Graph Construction and Adaptive Retrieval
30mFusing NLP and Graph: Building a Conversational Agent for Enriched Financial Data
32mBuilding an Advanced Knowledge Assistant
30mHow to Build the Next Wave of Graph-Based Agentic AI