Data Mart and shared profiles

Data Mart manages local tables and MCP logical datasets in one place, but they are read differently. File and ordinary connector imports are usually table snapshots. Text material supplies reference content. An MCP dataset defines an external business scope for on-demand reads without copying the entire service.

One connection, several datasets

One MCP connection can support separate projects, repositories or semantic models. The connection holds address and authentication settings; each dataset has its own purpose, scope, validation and field meanings. Open a dataset's details to inspect or edit that scope. Adding another dataset does not replace the first.

Select each required dataset explicitly for an agent. Siblings on the same connection are not automatically authorized. A metric dependency does not open all underlying data. Members and shared recipients use the agent's authorized sources, not dataset editing controls.

Draft, saved and verified

Source, capability or authorization changes can make old evidence stale. Follow the recheck guidance instead of relying on an old summary.

Inspect and correct shared profiles on demand

Open Profile in an MCP dataset's details to inspect its summary, field meanings, units, caveats and original verification time. Loading a profile does not require a new provider exploration every time Data opens. Owners and collaborators can explicitly save corrections to shared metadata. This is not publication of another caller's private exploration history.

Editing wording, meaning or units does not run a new query validation, replace the original verification time with the correction time or automatically change executable metric contracts. Return to onboarding checks for evidence of read capability. Related knowledge candidates still need separate review.

Compare history and recognize stale sources

Open profile history and select before/after versions. Compare summaries, fields, reviewed definitions and subsequent decisions. Check verification and correction times and stale-source warnings. Older entries may contain only partial knowledge evidence; a preserved baseline does not establish the original import or correction date.

History supports comparison and attribution, not a promised one-click rollback. To restore wording, check the current source and save an explicit correction. Managers separately decide whether confirmed shared knowledge should be retained, revised or withdrawn; a source change does not overwrite human decisions.

Related: MCP onboarding, Data quality, Data overview and Copilot, Knowledge.