Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.cube.dev/llms.txt

Use this file to discover all available pages before exploring further.

The Dashboard Agent is a conversational assistant docked as a panel on a published or embedded dashboard. Ask it questions in plain language about what’s on the dashboard — and it answers using live queries against your semantic model, not just the numbers already rendered on screen. Because it queries the model directly, the agent can go beyond what a chart shows. It can drill into a single chart, re-aggregate a metric, break a number down by a dimension, pull a longer time range than the chart displays, and summarize trends across the dashboard.
Dashboard Agent panel docked alongside a published dashboard
There are two different agent surfaces with different capabilities:
  • Dashboard Agent (this page) — read-only Q&A for viewers of a published or embedded dashboard.
  • Workbook Agent — full authoring inside a workbook: creating reports, building and editing dashboards, and running analysis.
Most of this page describes the read-only Dashboard Agent. Keep the distinction in mind — the published agent intentionally cannot author or change anything.

Opening the agent

Open a published dashboard and toggle the agent panel from the dashboard header to start a conversation. Use New Chat to start a fresh thread. The panel is docked to the right side of the dashboard and is resizable — drag its edge to give the conversation more or less room. On narrow screens it switches to a compact mode so it stays usable next to the dashboard.

Focusing on a specific chart

Chart-level follow-ups and attached-chart chips are rolling out. If you don’t see the Ask a follow up question action on a chart’s menu yet, it hasn’t reached your deployment.
To ask about one chart in particular, open that chart tile’s menu () and choose Ask a follow up question. This opens the agent with that chart attached as context, so your question is scoped to that chart’s data. Attached charts appear as removable chips in the chat input. To drop a chart’s focus, remove its chip. You can attach more than one chart to ask about them together. When a chart is attached, the agent receives a read-only snapshot of it — the underlying SQL, the data, and the chart spec — along with the dashboard’s current filter and time-grain state. It uses this as context; it does not modify any of it.

What the Dashboard Agent can and can’t do

The published Dashboard Agent is read-only by design. It can:
  • Answer questions about what’s on the dashboard from the dashboard’s own data
  • Run read-only queries against the semantic model to go deeper — drill down, break a metric down by a dimension, or pull a longer time range than a chart shows
  • Explain how a metric or dimension is defined
  • Look up the values available for a dimension
  • Read the contents of attachments you add to the chat
It can’t:
  • Edit the dashboard, or add/remove widgets
  • Change filters or the time grain (see Limitations)
  • Create reports or workbooks
  • Change the data model
If you need any of those, use the Workbook Agent instead.

Attachments

You can attach uploaded files to the conversation, and the agent can read their contents — text, PDF, ZIP, and image files are supported.
Pasting an image directly from your clipboard into the chat may also be supported. Confirm this in your deployment before relying on it.

Limitations

  • Filters are read-only. The agent can tell you which filters and time grain are currently applied and reason about them, but it cannot change them for you yet. This is the most common expectation gap — if you ask the agent to “filter the dashboard to last quarter,” it will explain the current state rather than apply the change. This is actively being worked on.
  • No authoring on published dashboards. The published Dashboard Agent cannot create reports or build dashboards. Those live in the Workbook Agent.
  • Report search covers published reports only. When the agent searches for existing reports, it sees published reports.