Skip to content
Documentation
Cookbook
Capabilities: RAG & Retrieval
Text to SQL

Text to SQL

Use this pattern when users ask analytical questions in natural language, but your real system boundary is still a database.

The difficulty is rarely only SQL generation. The real work is grounding the model in schema context, constraining execution, and checking whether the answer is actually correct.

What to focus on

  • Getting reliable SQL generation (schema grounding + constraints)
  • Evaluating correctness beyond “looks right”
  • Defensive execution (validation, limits, safety)

When it works well

  • The question maps to structured data already stored in tables
  • You can provide enough schema context without exposing sensitive data
  • Safety controls matter as much as answer quality

If you want to reproduce it locally

After your local Cookbook environment is ready, you can validate the matching example with:

make test-notebooks NOTEBOOK=capabilities/text_to_sql/guide.ipynb
Knowledge GraphSummarization