On this page
Ask questions about an uploaded dataset
Ask MiniUp to create an app from your spreadsheet or CSV. The published app keeps its hosted Parquet data and DuckDB analytics. MiniUp also prepares two read-only capabilities automatically, using the dataset’s actual fields:
- Query: search text, filter records, choose fields, sort, and page results. Date and numeric comparisons support ranges.
- Summarize: count records or calculate sums, averages, minimums, and maximums, optionally grouped by a field. Results can be ranked by the aggregate value.
For example, ask “Find PM Construction records,” “Show records issued after January 1, 2026,” or “How many records does each contractor have?” The relevant fields must exist in your dataset; no field names are hard-coded.
There is no extra setup tool, editable companion Table, or backend choice to learn. Browser agents and Remote agents become ready automatically when eligible actions are enabled. Advanced controls, testing, and activity remain available in Agent Actions.
Access and refresh
Dataset actions are always read-only, including for owners and editors. They use the app’s existing access and membership rules. Protected apps remain protected. Remote access to a protected app requires current membership. Storage keys, signed URLs, and credentials are never part of the action definition or response metadata.
Replacing a dataset refreshes its action fields. Disabling or deleting it immediately stops execution. Manual names, descriptions, disabled actions, removed capabilities, and protocol preferences survive refreshes and publication. Inputs are checked against the live schema; unknown fields and raw SQL are rejected.
Query limits
Each query returns at most 100 rows or groups (25 by default), supports at most 12 filters, and permits a row offset up to 10,000. Numeric fields support numeric aggregates; date fields support minimum and maximum. Count does not require an aggregate field. Grouped results sort by aggregate value, descending by default.
The initial server query boundary is 16 MiB of Parquet, 50,000 rows, 64 fields, and 64 MiB of declared/decoded column data, with a 10-second deadline covering download and execution. Nested or repeated Parquet schemas are unsupported. Geometry-only datasets have no queryable scalar fields. Cells are limited to 8 KiB, a decompressed page to 4 MiB, groups to 1,000, retained sorting data to 4 MiB, and a response to 256 KiB. Exceeding a limit returns an error rather than a partial summary.
Files stream to temporary disk; projected columns are decoded in a worker with a bounded heap. Two queries may run concurrently per server instance. Existing Agent Action request limits also apply. Queries currently scan the bounded file; this is not a large-dataset analytics engine. Larger datasets continue to work in the human-facing DuckDB app, but their agent queries are unavailable. Narrowing returned fields or using summaries helps with response size, but does not bypass file or scan limits.