What We Learned Building LLM-Powered Text-to-SQL

"What We Learned Building LLM-Powered Text-to-SQL," presented by Fatma Ozcan, Principal Engineer at Systems Research @ Google. This workshop provides a technical overview of the challenges behind natural-language-to-SQL systems, how large language models generate accurate queries, and how frameworks like CHASE-SQL and contextual signals drive major improvements in performance.

What You'll Learn:

• Why text-to-SQL remains challenging in real-world, large-scale databases
• How LLMs translate natural language into structured, executable SQL queries
• An inside look at CHASE-SQL, a multi-agent framework for improved query generation
• Key findings from achieving strong results on benchmarks such as BIRD and Spider
• How contextual signals—examples, hints, schema, and documentation—boost accuracy and reliability

Fatma Ozcan is a Principal Engineer at Systems Research @ Google. Previously, she served as a Distinguished Research Staff Member and senior manager at IBM Almaden Research Center. Her research focuses on platforms and infrastructure for large-scale data analysis, query processing and optimization for semi-structured data, and democratizing analytics through natural-language querying and conversational interfaces.