[Digital Today reporter Chi-gyu Hwang (황치규)] Database and analytics companies are rushing to release text-to-SQL tools that query databases using natural language. Experts are warning that caution is needed when non-specialists use these tools on their own.

According to a recent report by The Register, AWS recently disclosed a text-to-SQL solution using its Bedrock platform. Snowflake introduced Cortex Analyst, which links business language such as "last month's revenue" to technical terms that databases understand. MongoDB also released its own natural-language query API based on LangChain.

Nick Koudas (닉 쿠다스), a professor in the University of Toronto's computer science department, warned, "The problem is that when business users use it, the system can generate queries that are grammatically correct but different from the intent." Current text-to-SQL system accuracy is about 80 percent, below human experts at about 93 percent. Most companies have proprietary data and internal jargon, making it difficult for LLMs to understand without additional training, he said.

A bigger risk is a query that is grammatically correct but wrong in meaning. Grammar errors are immediately obvious because the query does not run, but a query that is wrong in meaning can produce irrelevant results without looking like an error, The Register reported.

Koudas said that for now, a process in which non-specialists write SQL queries in natural language needs experts who understand SQL and can verify the results to take part.

Researchers are also exploring ways to raise accuracy by having LLMs ask users follow-up questions. "Natural language is ambiguous and full of nuance," The Register reported, describing an approach in which "when an LLM encounters an uncertain token, it asks the user whether the expression means A or B" and completes the query together.

Koudas said text-to-SQL is a tool to increase developer productivity, not a technology that replaces developers.