Prepare your data

What an agent can answer depends on the queryable data and the definitions you approve. This page explains how to add local files, connect upstream data and keep its snapshot current.

The Data page

Data (/data) separates sources from the query layer:

• Connections manages workspace connections for Google Sheets, Feishu Bitable, S3 and Snowflake.
• Files lists uploaded and imported datasets, their sources, structure, quality, refresh status and dependencies.
• Data Mart shows the query engine's tables, sample rows and semantic information. Agents query data through this layer.
• Documents are reference material for agents, not aggregatable tables.

A new dataset normally appears in Files first and then in Data Mart after ingestion. Uploading a dataset does not expose it to every agent automatically.

Add local data

File upload supports CSV, TSV, XLSX and JSON. For a multi-sheet XLSX, choose the sheet to import; the workbook is never imported wholesale without a selection. You can also provide a public, directly reachable CSV or JSON URL over HTTPS. A web page, login-protected address or HTML response is not a data file. The platform downloads and stores a snapshot, records the source URL and can fetch it again later.

One import:

1. writes the rows to the query engine and stores the column structure and samples;
2. builds a semantic profile for each column (type, date/category/measure/ identifier role and statistics); and
3. drafts metrics, dimensions and starter views for your review.

Review proposed metrics and approve only definitions that match your business rules. Edit names, descriptions or formulas where needed. Approved metrics enter the metric catalog; governed agents use those defined metrics rather than ambiguous column names.

Connect upstream data

Connector imports are snapshots of upstream data. You can refresh them from the same source later:

S3 can also provide an external (zero-copy) dataset when your plan allows it. External mode accepts Parquet only: the data stays in your bucket and the platform stores metadata and reads the object at query time. It is not a local copy, so local cleaning, append and re-upload are unavailable. Refresh checks remote reachability, schema and metadata rather than copying rows.

HTTP/API, SMTP and webhook connections are tools, delivery channels or inbound verification capabilities; they do not become dataset connectors merely because you configured them. Credentials are encrypted and never shown again. Connections belong to the workspace and can be reused by multiple datasets.

Keep data fresh

Datasets with a live connector or URL source support Refresh now. Refresh re-pulls the source and replaces the dataset's existing rows. A local upload has no source to re-pull.

Automatic schedules are:

• off — no automatic refresh; refresh manually when allowed;
• hourly — every hour;
• daily — every day; and
• weekly — every week.

Only datasets with a live source can have a schedule, and only workspace owners or admins can configure schedules or run a manual refresh. At ask time, the system may still refresh a live source when its refresh window is due or when the request explicitly asks for a forced refresh. If refresh fails, the last successful snapshot remains available for answers and the Data page shows the failure.

Normal connector and URL refreshes require the incoming columns to match the existing structure exactly. Added, removed or renamed columns are reported as schema drift instead of silently replacing a table that metrics depend on. If you really intend to adopt the new structure, explicitly confirm a forced re-import; metrics that use removed or renamed columns may then stop working.

Re-importing a local file offers two modes:

• Replace — replace the existing rows with the new file.
• Append — add the new rows after the existing rows.

Both modes check the column structure and use versioned writes with rollback protection. External zero-copy datasets reject local re-import; fix connector data upstream and refresh instead.

What agents can read

Metrics link to their datasets through their formulas and can be grouped into catalogs or subject areas. In the agent wizard, choose catalogs or an explicit data scope. The agent then receives the corresponding datasets and certified metrics, while drafts and deprecated metrics do not silently enter its governed scope. A catalog subscription follows later catalog membership changes; an individual metric selection remains explicit.

Uploading or connecting data does not automatically grant every agent access. Workspace settings also offer Explore, Governed and Hybrid data modes: Explore permits raw-data tools, Governed answers from governed metrics and hides raw SQL/file tools, and Hybrid prefers governed metrics but may fall back to raw data with an ungoverned label. Separate settings control raw SQL, record export and conversational data-upload capabilities; governed detail reads can remain available. Governance does not replace column descriptions or metric review: write down the business definition first, then grant the needed scope.

Next

• Data quality and the semantic layer
• Create an agent