VoiceGraph
A knowledge graph you question out loud. Built solo, overnight, at the Build with AI hackathon run by Google at NYU.
What it is
You talk; a Gemini Live agent decides how to search a Neo4j knowledge graph, answers out loud, and a React and Three.js scene redraws the part of the graph it used.
How it works
- Eight retrieval strategies the agent chooses between: semantic search, shortest path, Leiden community detection, and LLM-generated Cypher among them.
- Agent tools and two sub-agents built on Google's Agent Development Kit, with a separate set of voice tools for the live audio loop.
- FastAPI and WebSocket backend, deployed on Cloud Run. About 12,000 lines, written in one night.
Status
The hosted demo is offline right now. The code is on GitHub.