This text discusses building knowledge graph agents with LlamaIndex Workflows for retrieving data from a Neo4j graph database using retrieval-augmented generation (RAG). The authors explore various workflow architectures, including Text2Cypher, which converts natural language questions into Cypher queries. Three workflows are implemented: Naive Text2Cypher Flow, Naive Text2Cypher with Retry Flow, and Iterative Planner Flow. The Naive Text2Cypher with Retry Flow improves accuracy by allowing the system to retry and correct invalid Cypher statements. The Iterative Planner Flow introduces a more sophisticated approach by breaking down complex questions into smaller steps and validating each subquery before execution. Benchmarking results show that adding a retry mechanism and final verification step significantly boosts answer relevancy. The authors discuss learnings, production-ready steps, and challenges such as handling null values and schema limitations. They emphasize the importance of establishing a solid benchmark to evaluate and compare different architectures effectively and encourage experimentation with LlamaIndex Workflows.