Last month, Tower hosted a technical seminar in our India office titled “Beyond Prompting: Building Reliable Text-to-SQL Systems with LLMs.” The session brought together our core engineering and quantitative trading teams to explore better ways to build reliable language interfaces to databases.

We were glad to welcome Prof. Sunita Sarawagi from the Department of Computer Science & Engineering at IIT Bombay as our guest speaker. Prof. Sarawagi led a discussion on the challenges of handling ambiguity, adaptation, and uncertainty when using large language models (LLMs) to query complex database structures.

While standard LLMs offer promising natural language capabilities, base models often struggle in enterprise settings, resulting in lower reliability due to intricate schemas and domain-specific linguistic nuances. During the talk, Prof. Sarawagi outlined actionable strategies to overcome these limitations, including:

• Schema Subsetting & Profiling: Utilizing LLM-driven schema sub-selection and logical semantic layers to simplify and accurately map enterprise database structures.
• Verified Query Templates: Implementing structured, pre-verified templates to bypass slow, token-by-token generation in favor of faster, deterministic execution.
• Ambiguity Resolution: Leveraging plan-based sketching and interactive feedback loops to clarify complex user intent before execution.
• Multi-SQL Consensus & Reinforcement Learning: Applying multi-plan consensus mechanisms to optimize accuracy and performance across heterogeneous data sources.

In the fast-paced world of quantitative trading, immediate data accessibility, system speed, and absolute precision are paramount. Technical exchanges like this seminar form a key part of Tower’s internal learning culture, ensuring our systems and engineering teams remain at the forefront of AI research, high-performance computing, and data architecture.

Thank you to Prof. Sunita Sarawagi for sharing her valuable research and expertise with our teams, and to everyone who helped make the session a success.