GDS Agent for Graph Algorithmic Reasoning
summary
An agent that exposes Neo4j Graph Data Science algorithms over an MCP server runs locally and projects subgraphs, inspects schema, chooses algorithms and streams real GDS results back to the LLM. Demo uses a London Underground graph with 302 stations and 812 edges and runs yen k-shortest paths, centrality, components and community detection. Benchmarks over 45 implicit-algorithm questions show ~66% exact answers on the London data, up to ~85% with a strong model, but performance drops to ~50–60% on a heterogeneous Game of Thrones graph.
AI-written summary · about
chapters (8)
- 0:00Introduction to GDS agents
- 1:15Graph data science basics
- 3:28Agentic framework concepts
- 5:30Why build GDS agents
- 7:00Practical agent demonstration
- 16:00Benchmarks and performance
- 19:48Summary and future plans
- 21:24Advanced tasks and closing
related talks
16mGraph Traversal Agent for Code Debugging
08mHarness Graph Power With Agentic AI: An Insights Discovery Use Case
01h 00mGraph Mastery to AI
30mAgent Reasoning with Graph World Models
30mAgentic RAG with Knowledge Graphs Explained
36mAgentic GraphRAG: AI’s Logical Edge
32mMulti-Agent Query Generation w/ LLMs on Complex Knowledge Graphs
31mThe Searchable Grid: Building a GraphRAG-Powered Digital Twin for Real-Time Insights