Problem
Agents answer data questions fluently, but often with the wrong number. In the speakers' environment agent-driven queries are expected to outnumber human and job queries by an order of magnitude within two to three months. Asked "is that the right number?", an agent sometimes admits it got it wrong or made it up. It must map business concepts to data, know their definitions, and find the tables, joins and expressions. All it sees is a table like fct_revenue_daily, and table comments and lineage rarely give 95%+ reliability. Definitions also drift faster than people can maintain them.
Example from the article: "What is our revenue per active customer in Q3?" requires the agent to know what revenue and "active customer" mean, whether Q3 is fiscal, and which tables to join.
Solution in brief
Put business knowledge into a governed layer the agent reads: metric views for calculations, a glossary for concepts, domains to partition context. An inferred ontology on top covers what nobody modeled by hand. Metric views are computed at query time and routed to materializations for speed.
Questions this case answers
- How to make an analytics agent return the right number for business concepts by putting calculations and definitions in one governed layer
- How to compare periods and grains in a reusable metric definition and keep the layer fast when many pre-aggregated tables exist
- How to make an analytics agent answer business questions with the right number instead of a confident guess
- How to define metrics and business concepts once so agents, BI tools and SQL use the same definitions
Tools and methods
- Unity Catalog metric views — reusable measures and dimensions over star and snowflake schemas
- Glossary and domains — governed concept definitions linked to assets; domains partition context
- Genie Ontology — inferred, permission-aware graph
- Genie Code — agent that drafts metric views and their YAML
- Materialization with query rewrite — queries routed to pre-computed tables
Google or a one-time email code. No password.