Talk to Your Graph: A Practical Guide to Building a Dual-LLM Q&A System for a Fashion Graph

Session Track: AI Engineering

Session Time: 07:15 am - 07:30 am November 6

Session description

Connecting LLMs to structured knowledge graphs often leads to unreliable queries and erodes trust in AI-driven analytics. How can developers build conversational systems that are not just powerful but also predictable and accurate? In this session, Vaskya and Celine will show you their practical end-to-end solution. They will guide you through a system, built from scratch with Python, that transforms a raw fashion dataset into a fully interactive knowledge graph using Neo4j and Google Gemini. The talk will first unveil a novel technique where an LLM acts as an "AI Architect" to automatically design the graph schema. Then, you will get a deep dive into the core of their Dual-LLM Q&A architecture, where one AI agent translates questions into high-accuracy Cypher using "Golden Rule" constraints, and a second agent narrates the data back into human-readable insights. By the end of this session, you will learn how to: - Architect a reliable Dual-LLM chain for separating query generation from summarization - Implement constraint-based prompting to enforce guardrails and dramatically improve Cypher accuracy - Automate the initial, time-consuming graph modeling process using an LLM - Build a complete, conversational application by integrating Neo4j and Google Gemini