Developing, Sharing, and Maintaining AI/BI Genie Spaces
What Genie is
An AI/BI Genie space lets business users ask questions in natural language and receive SQL-backed answers, tables, and charts. Genie is not a general chatbot: it reasons over a curated semantic context you provide, so answer quality is a function of curation quality.
What goes into a space
- Tables, views, and metric views - the trusted assets Genie may query.
- Instructions - plain-language guidance on business rules, definitions, and how to join tables.
- Sample/benchmark questions - examples that teach Genie the expected phrasing and correct SQL; used to measure accuracy.
- Comments and column descriptions on the underlying tables, which Genie reads.
- Example SQL queries attached to assets for tricky calculations.
Sharing and maintaining
Share a space with users or groups and grant them access to the warehouse plus SELECT on the underlying objects. Monitor usage and accuracy through feedback (thumbs up/down), the benchmark questions, and audit logs. Maintenance is ongoing: when a table changes or a metric definition shifts, update instructions and re-validate benchmarks.
Traps
- Pointing Genie at raw bronze tables with cryptic column names and expecting good answers.
- Granting space access but not warehouse or table privileges - users see errors, not answers.
- Treating sample questions as cosmetic; they materially affect accuracy.
- Forgetting to update the space after schema changes, producing silently wrong SQL.
Rule of thumb
Curate narrow and well-described. Five well-documented gold tables with clear column comments, a metric view, and ten benchmark questions outperform fifty undocumented tables every time.
Worked example
A finance team asks "revenue last quarter by region." Success requires: a gold_revenue table with a revenue_usd column comment defining the metric, an instruction that fiscal_quarter differs from calendar quarter, and a benchmark question covering it. Then Genie returns a defensible answer.