Context
The problem
Vector search is great at finding similar text and bad at questions that hop across relationships, like “who supplies the companies this group owns?”. I wanted answers that follow real connections instead of guessing.
Approach
How I built it
Gemini pulls people, companies and the links between them out of the text. Everything is checked against a schema before it's saved to Neo4j.
To answer a question, the agent writes a Cypher query, runs it on the graph and explains the result in plain English.
Generated queries are read-only, and every step has validation and error handling, so a bad query can't damage the graph.
Result
What came out of it
Answers that follow real relationships
Read-only queries by design
Errors caught at every step