Grounding & Citations
How It Works
Before any query, the agent reads the semantic model: which marts answer the question, at what grain, which metric owns the calculation, and which rules constrain a valid answer. Then it queries and checks the result before reporting.
Citations
Definitions used appear as clickable links in the answer — the metric or dataset itself, not a footnote.
They take the form [Revenue](ekai-ref:dp/…) and resolve in the ekai UI. In exported markdown they are inert text.
Readers cannot audit the SQL, but they can open the definition behind a number in one click and see whether it is the rule they meant.
Where a metric exists, its stored formula is used rather than re-derived — so two people asking the same question in different words get the same number.
Worked Examples
| Ask | What comes back |
|---|---|
| "How many customers have never ordered?" | One sentence with the figure and its share of the base. No chart. |
| "Rank nations by customer count." | One ranked bar and a line saying where it turns. |
| "Show order value in $25k bands." | One distribution chart, with the tail called out. |
| "Why is on-time delivery so low?" | The rate, then the one split that moves it, then the caveat that order-level and line-level rules are different measures. |
| "Demonstrate the data ranges via graphs." | A report canvas: coverage, numeric ranges, categorical spreads, and what the model deliberately does not carry. |
| "What does on time mean here?" | The definition as the business wrote it, with a link to it. No query needed. |
A good question names the population, the period and the measure. "On-time rate by market segment for 1997" needs no clarification; "how are we doing" needs three.