Walk through an example end-to-end journey of building a Databricks Genie space with benchmarks to develop a trustworthy system.

databricks

Building production-ready Genie spaces requires systematic validation through benchmarks rather than subjective assessment. The process involves improving Unity Catalog metadata (table/column names and descriptions), defining explicit relationships through primary and foreign keys, enabling value dictionaries for accurate filtering, adding example SQL queries for custom metrics, and using text-based instructions for business rules. Starting with 0% accuracy on 13 benchmark questions, each iteration addressed specific gaps: metadata cleanup enabled single-table queries, relationship definitions fixed multi-table joins, value sampling improved filtering, example queries taught business logic, and instructions handled edge cases. The systematic approach transforms trust-building from reactive problem-solving to proactive validation, ensuring users receive accurate results before deployment.

How to Build Production-Ready Genie Spaces, and Build Trust Along the Way

The Journey: Developing from Baseline to Production

What We Learned: The Impact of Benchmark Driven Development