Import & sources

Datasets are your raw tables. Every one is versioned, and every one carries sources — citations for where the data came from — so any number in the workspace can be traced back to a file or a URL.

The import wizard

Datasets → Import data walks you through three steps:

1. Upload a CSV, JSON, or Parquet file (gzipped files work too).
2. Preview how it parses. For CSVs you can adjust the delimiter, header row, and rows to skip until the preview looks right — nothing lands until you confirm.
3. Import. The file becomes a named dataset backed by an immutable, versioned table.

Besides creating a new dataset, the wizard can target an existing one:

• Append adds the new rows to the current data.
• Replace lands a new version in its place — history is kept, so nothing is lost.

The dataset page

Open any dataset to work with it:

• Preview shows the current rows.
• Schema lists the columns — and here the metadata is editable: set each column's semantic type, unit, scale, and description. This is what makes a column called amt legible as "funding amount, EUR millions" to teammates and to agents reading the workspace.

Sources: why citations matter

A table of numbers with no origin is a liability — you cannot defend it in a meeting, and neither can an agent reasoning from it. Sources fix that: on the dataset page, add a citation for each place the data came from (a URL, a report, a file) plus any judgement applied while preparing it. Datasets that agents load arrive with sources already attached; datasets you import deserve the same care.

Faster with an agent — ask the agent to "research <topic>, clean the results, and load them as a dataset with sources." It gathers the data, normalizes it into a CSV, uploads it, documents the columns, and cites every origin — one conversation instead of an afternoon of copy-paste.