The article outlines an architectural framework for integrating Large Language Models (LLMs) with relational databases, moving beyond the typical read-only systems to enable both reading and writing capabilities. By utilizing a Knowledge Graph (KG) as an intermediary, the framework addresses the challenges of synchronization between the KG and the relational database. The author evaluates different architectural patterns, highlighting the importance of determining which system acts as the "source of truth" and how data changes are managed. Three primary architectural patterns are discussed. The first positions the KG as the main operational database, with the relational database serving as a downstream replica. This setup offers immediate consistency but requires a significant cultural shift for organizations. The second approach, a hybrid model based on Command Query Responsibility Segregation (CQRS), separates the reading and writing processes. In this model, the relational database handles transactions while the KG manages complex queries, optimizing performance for both tasks. The article provides practical guidance on implementing these designs, including best practices for data modeling, prompt engineering, and security measures. Specific strategies are outlined for generating commands, such as using Cypher for database interactions and ensuring safety in data operations. This comprehensive approach aims to create more interactive and capable chatbots that can manipulate relational data effectively, ultimately enhancing user experience in data-driven applications.