Conscious Engines — Text-to-SQL

Generated SQL

—

Idle

The generated query will appear here.

Model Reasoning

Details

Modelsql-velox
Parameters7B
Latency—
Prompt tokens—
Completion tokens—

About

sql-velox is a compact 7B reasoning model specialised for text-to-SQL. Reinforcement-trained on query execution, it matches far larger general models at turning natural language into SQL while staying small enough to self-host. Served behind an OpenAI-compatible API on a single GPU.

It never sees your data — only the schema (the table and column definitions) goes into the prompt. The model writes a query; running that query against the database is a separate step, so no rows ever leave your system.

~68% BIRD execution accuracy — best in its size class

12/12 on our neutral in-house benchmark

And this is the base model out of the box. With our task-specific tuning, custom harness and query-verification layer, we typically lift execution accuracy a further ~8–10 points on a customer’s own schemas.

Raw Response

{}

Text⁠-⁠to⁠-⁠SQL

Generated SQL

Result