The situation
Retail questions are relationship questions: which products, in which stores, for which customer segments, over which periods. Flat tables and a generic chatbot give plausible answers that fall apart under follow-up.
What we built
A multi-agent question answering and intelligence system on a Neo4j knowledge graph. Separate agents handle graph query generation in Cypher, financial insight, forecasting and chart generation. Visualizations are produced programmatically as React charts for each question, rather than chosen from a fixed set.
How it is engineered
- The graph is the shared source. Entities and relationships are modelled explicitly, so an agent can traverse a business question instead of guessing across a flat table.
- Narrow agents, explicit hand-offs. Each agent has one responsibility, which makes each one testable and makes failures easy to locate. The agents are built with LangChain Deep Agents.
- Deployed on Google Cloud, with Redis alongside the graph and the agents.
Why it matters
The same pattern, agents reasoning over a graph with narrow responsibilities, suits any domain where relationships carry most of the meaning, such as provider and expert networks in life sciences.