Ask questions

Questions become governed Cube queries — never raw SQL. Every step shows up below, including the ones the gate rejects.

How a question is answered

The agent never writes SQL. It searches the model, writes a small Cube JSON query naming members from it, and that query passes four checks before it touches your warehouse:

Check	What	Cost
1. Shape	Is it a valid Cube query?	free
2. Meaning	Do the names exist? Does each filter fit its type? Is the time grain available?	free
3. Compile	Can Cube turn it into SQL? Nothing runs yet	one small call
4. Run	Run it, capped at 5,000 rows	warehouse

A failed check comes back to the agent as a message it can act on — "orders.cnt does not exist. Closest matches: orders.count" — and it corrects the query. You see the rejection and the retry.

What the transcript shows

Block	What you see
Thinking	The model's reasoning summary. Collapsible; Thought for Xs when done
Search	What it looked up in the model, and how many members matched
Query	A plain description of the query, rows and time. Expand for the Cube query, the SQL Cube made, warnings and errors, each with Copy
Result	A sortable table (N of M rows, Show all, CSV download) and a chart when one fits
Python	The script, its approval, its live output, and any files it wrote
Answer	The written answer, with Copy answer

A chart is chosen from the result's shape. No chart is a normal answer: some results are only a table. You can ask for a different chart ("show that as a stacked bar").

Sorting a result sorts the rows loaded, not every row: the note says Sorted over the N rows loaded, not all M when it matters.

Files and Python

Attach a file with + or drag it onto the window. It goes to the sandbox, which reads its shape — rows, columns, types — for the agent. Up to 100 MB, any type.

The agent can run Python in the sandbox for what a query cannot express: ratios, period-over-period changes, shares, joining your uploaded file to query results. It moves a query's stored rows there with export_result.

Each script is shown before it runs: Run this script?

127.0.0.1:8420/#chat

What share of each region's revenue came from Produce, this week vs last?

✓Total revenue by region, category, day · last 14 days392 rows · 1.1s ▸

✓sales_14d.csv0.4s

⏸produce_share.pywaiting for you ▾

import pandas as pd df = pd.read_csv("uploads/sales_14d.csv") day = pd.to_datetime(df["daily_sales.sale_date.day"]) df["week"] = (day > day.max() - pd.Timedelta(days=7)).map({True: "this", False: "last"}) keys = ["daily_sales.dim_stores_region", "week"] total = df.groupby(keys)["daily_sales.total_revenue"].sum() produce = df[df["daily_sales.dim_items_category"] == "Produce"] share = produce.groupby(keys)["daily_sales.total_revenue"].sum() / total print(share.unstack().round(3))

Run this script?RunAlways in this chatDecline

A script waiting for approval. The whole script is shown; nothing runs until you choose.

• Run — run it once;
• Always in this chat — stop asking in this chat until the portal restarts;
• Decline — the agent is told you said no.

Images a script makes show inline; CSVs and text show a preview and a Download link.

Controls

• Enter sends; Shift + Enter is a new line.
• Stop ends the agent's turn. A script running in the sandbox is killed.
• Long chats are summarised to fit the model's context; a notice says when.
• Rename a chat by double-clicking it; delete with ×, then Delete.

Saving to a dashboard

There is no button for this. Ask in the chat: "save that to the Sales dashboard". The agent adds a tile with the query — see Dashboards.