Data
Give your Project something to query. There are two ways to do it, and which one you want depends on how settled the questions are.
Audience: builders
Two ways in
Attach a data source and let the agent write the queries. Your administrator connects the warehouse or database once; you attach it to this Project. From then on the agent reads the schema and composes SQL to answer what you ask. There is nothing to model and nothing to maintain.
Or build a Data Collection. A collection sits on top of one or more sources and adds the parts a raw connection cannot give you: your business vocabulary, a catalog the agent reads instead of guessing from column names, and verified queries your team has approved. You attach the finished collection to the Project the same way.
Start with the data source. Move to a collection when you hit one of these:
| Reach for a Data Collection when | Why a raw source is not enough |
|---|---|
| The same question gets asked constantly and the answer must not drift | A collection can publish a verified query: approved SQL, same answer every time |
| Your tables do not use the words your team uses | A collection defines what Customer, Invoice, or Days Payable Outstanding mean, so questions are asked in your language |
| The answer spans several systems | A collection blends multiple sources behind one queryable surface |
| Someone has to sign off on the numbers | A collection's answers carry lineage, and the questions are answered against a reviewed catalog |
| You want the data available in other AI tools | A collection publishes an MCP endpoint; a raw source does not |
You are not choosing once and for all. A Project can have both. Attaching a data source directly is the fastest way to find out what your team actually asks, and that is the best possible input to a collection when you build one. See Data Collections for the authoring side.
Where both live
Open your Project in Build. The Data group in the Context tree holds two rows: Data sources and Data Collections. Each lists what is attached here, with health visible on the row.
Attach a data source
Open Data sources
The row lists the organization-managed sources available to you. If it is empty, nothing has been connected for your workspace yet: your administrator sets these up under Connections.
Attach the one you need
The source moves into the attached list, and health and discovery for it stay visible on the row.
Ask a question
In a conversation, ask for what you want in plain language. The agent reads the schema and composes the query. If it gets a join wrong or picks the wrong column, tell it, and save the correction as a memory so it holds next time.
Attach a Data Collection
Open Data Collections and click Attach
A picker lists every collection the workspace has published that is not already attached here. Search by name or description.
Pick a collection
The picker closes and the row moves into the attached list, showing name, semantic summary, source count, and build status: Healthy, Populating, or Failed.
If the picker is empty, either every workspace collection is attached here already, or nobody has published one yet. Authors publish from the collection's own page; see Create a Data Collection.
Attached collections are available to every conversation and every Skill in the Project.
What attachment lights up
- In a conversation. The collection appears in the
@picker. Mention it to ground the next answer in its catalog. - In a Workflow. Reference the collection by name and it becomes a
data_domain.intentstep. The agent picks the right tables and columns from the catalog. - In audit. Every query the Project runs writes a ledger entry, joined to the conversation or work item that triggered it.
A Failed chip means the collection's last build did not complete. Open the row drawer for the detail and a link to the collection's own page.
Refresh and detach
A collection's catalog rebuilds on the collection's own page, not from inside the Project. The attached row reflects the latest published build, picked up on Project reload. If you need fresher data, work with the collection's author.
To detach, open the row's detail drawer and use Detach in the footer. The collection itself stays in the workspace, and other Projects keep their attachments.
Attached time versus build time. The row shows when you attached the collection here. The semantic build phase shows when the collection itself last rebuilt. The two are independent, and detaching and reattaching does not trigger a rebuild.
What audit captures
Every attach and detach writes a ledger entry. So does every query the Project runs, with the collection or source name, the intent, the tables involved, and the result shape. Open /audit and filter by the Project to walk the trail.