A natural language to SQL query engine that translates plain English questions into executable SQL, evaluated against BIRD-SQL bnechmark.
You ask a question in plain English:
"Who made the most sales last quarter?"
The engine translates it to SQL, runs it against the database, and returns the answer — no SQL knowledge required.
The validation loop is the agentic part. Unlike a simple pipeline where each step is predetermined, the loop:
- Acts — runs the generated SQL
- Observes — checks whether it succeeded or failed, and how it failed
- Decides — chooses between returning the result, fixing a syntax error, or rethinking a semantic mismatch
- Reacts — generates a new attempt with the error as additional context No human is involved between steps 2 and 3. The system decides what to do next based on what it observed. That is what distinguishes an agent from a fixed workflow.
Evaluated against the BIRD-SQL benchmark — 10,000+ natural language / SQL pairs across 11 real-world databases.
- Python 3.11+
- SQLite (via Python's built-in
sqlite3) - Meta llama-3.3-70b-versatile
MIT