Feyn AI’s new SQRL family can probe a database before generating SQL, boosting accuracy for ambiguous queries.

Feyn AI has unveiled SQRL, a novel family of text‑to‑SQL models that can examine a target database before crafting a query, dramatically improving accuracy on ambiguous natural‑language requests.

What Is SQRL?

SQRL (SQL Query Retrieval Language) is a suite of large language models trained to first retrieve schema and sample data from a database, then generate SQL statements that align with the observed structure. This two‑step approach helps the model avoid common pitfalls such as referencing non‑existent tables or misinterpreting column types.

How the Inspection Works

When a user submits a question like “Show me the top‑selling products last month,” SQRL first issues a lightweight introspection query that returns the relevant tables, column names, and data samples. The model then uses this context to produce a precise SELECT statement, reducing hallucinations that plague generic text‑to‑SQL systems.

Performance Gains

In internal benchmarks, SQRL achieved up to a 20% boost in exact‑match accuracy compared with baseline models that generate SQL without schema awareness. The improvement is most pronounced on complex schemas where column names are ambiguous or where multiple tables share similar attributes.

  • Improved handling of ambiguous column names
  • Reduced generation of syntactically invalid queries
  • Better alignment with database constraints

Potential Use Cases

Enterprises can embed SQRL into analytics dashboards, chatbot assistants, and self‑service reporting tools, allowing non‑technical staff to ask natural‑language questions while the system ensures the resulting SQL is safe and accurate.

“By letting the model see the data first, we bridge the gap between language understanding and database reality,” said Dr. Maya Patel, lead researcher at Feyn AI.

The SQRL family is available via Feyn AI’s API platform, with plans for on‑premise deployment for organizations with strict data governance requirements.

For a detailed overview, see MarkTechPost coverage of Feyn AI’s SQRL launch.