Lightweight 'Ask your data' chat panel for dashboards

[cliftonc]

Summary

Add a lightweight chat UI component that can be embedded on dashboards, allowing non-technical users to ask questions about their data conversationally.

Context

This is a P1 gap identified in a competitive analysis against Omni. Omni has an "Ask AI" chat accessible from any dashboard. Drizby/drizzle-cube has powerful agentic notebooks with 6 tools (discover_cubes, get_cube_metadata, execute_query, add_portlet, add_markdown, save_as_dashboard), but they're oriented toward power users. Business users need a simpler entry point.

Implementation Notes

• The agent infrastructure and tools already exist in src/server/agent/tools.ts
• The full AgenticNotebook component exists with AgentChatPanel, ChatInput, ChatMessage
• This is a lighter-weight version — a collapsible chat panel/drawer on dashboards
• Should use the existing agent tools (primarily discover_cubes and execute_query)
• Could show inline chart results from add_portlet tool calls
• No need for full notebook canvas/blocks — just chat messages + inline results

Acceptance Criteria

• Chat panel component that can be added to any dashboard view
• Users can type natural language questions about their data
• AI uses existing agent tools to discover relevant cubes and execute queries
• Results shown inline (tables, simple charts)
• Collapsible/dismissable so it doesn't obstruct dashboard view

[cliftonc]

Sustain bot triage:

This looks well-scoped and not obviously duplicate with the existing dashboard/chart feature requests I checked. Keeping enhancement makes sense.

A few implementation notes that may help when this gets picked up:

• Reuse of the existing agent tool surface is a strong fit here; the main boundary to keep clean is UI shell vs. agent execution/session state.
• I’d recommend splitting delivery into 3 layers:
  1. dashboard chat drawer/panel container
  2. lightweight chat/session API that exposes only the safe dashboard-oriented tools
  3. inline result renderers for table/chart responses
• It may be worth constraining the first version to read-only actions (discover_cubes, get_cube_metadata, execute_query) and treating add_portlet as a follow-up once the UX is stable. That reduces the risk of the lightweight panel growing into a second notebook surface.
• For acceptance, consider adding one non-functional criterion around permission/security context propagation so dashboard queries always execute with the same row-level access as the surrounding view.

Overall: good candidate for a P1 roadmap item because the underlying agent primitives already exist and the gap is mostly packaging + UX.

[last-light]

Acknowledged — starting build cycle for #564. Checking for prior progress...