Lumian / Architecture walkthrough
How we handle large amounts of data at Lumian
Lumian works with orders, sales, inventory and advertising reports. The account, marketplace and reporting period define each question. Inventory also has its recorded snapshot date.
Architecture walkthrough only. No customer data is shown. This static page runs no live query or AI request.
Example question
“For the selected account and period, what do orders, sales, inventory and ads tell us?”
From stored reports to an explanation
| Step | What happens |
|---|---|
| Stored reports | Orders, sales, inventory and advertising data stay in the data layer. |
| Scoped queries | Read the selected account, marketplace and reporting window. Queries filter records, group them and calculate totals before the model explains them. |
| Bounded facts | Large results can become a compact model view with calculated summaries and selected rows. Row and text limits keep that view bounded. |
| AI explanation | The model explains the returned facts and their limits, rather than receiving the entire stored history. |
How Lumian Agent processes SQL
Account filters are applied to supported table reads before SQL validation. PostgreSQL filters and calculates; the model receives a bounded preview.
Keep exports separate from the model preview
Supported tabular tools retain returned rows separately for CSV export. An export contains the rows returned by its query; the source query can have its own limit.
Check freshness and coverage
Report reads check freshness and requested coverage. Where required, they seek confirmation of the exact reporting window and reread stored results. A report still being prepared remains pending. Source and reporting-window details stay attached to the result.
A limited result is not proof of complete coverage.