Dettagli sessione

Developing a dependable Text-to-SQL application using LLMs presents several practical challenges, from schema alignment and ambiguity handling to error recovery and user experience. While end-to-end approaches using a single LLM or agent may seem appealing, they often prove opaque and difficult to maintain in real-world scenarios. In this talk, I will share insights from my personal experience designing and implementing a modular Text-to-SQL system. Rather than relying on a monolithic agent, I adopted a task-specific approach by decomposing the problem into distinct stages. To orchestrate this architecture, I used the Haystack framework, which enabled flexible composition of the pipeline and simplified experimentation. I will present code examples of the modular pipeline and a simple agent, discussing the trade-offs encountered and the reasoning behind key design decisions. This talk will provide a practical perspective on how modular design can lead to more reliable and maintainable LLM-based applications, grounded in real-world development experience.