Data Insights — User Guide

Audience: Analysts and data owners who want conversational analytics: ask in natural language, get KPIs, charts, tables, and narrative insights in one place.

Overview

Data Insights (/data-insights) pairs a specialized Data Insights Agent with a live dashboard that updates as you chat. Unlike generic chat, this view is optimized for metrics, plots, and tabular results you can resize, inspect, and reuse.

Typical outcomes:

  • Executive-style KPIs with optional trend arrows
  • Charts (bar, line, pie, etc.—whatever the agent and renderer support)
  • Data tables for drill-down rows
  • Narrative bullets summarizing what matters
  • Suggested next charts you can click to generate

Layout: visualization + chat

Left: visualization area

Empty state — Prompts you to start the conversation.
Loading — “Generating visualization” style feedback while the first artifacts arrive.

When data exists, sections can include:

Block Description
KPI row Cards with title, value, description, icon (and color), optional trend (up/down with value and color).
Charts grid Responsive grid of chart cards. Each card has a title and actions to copy or view the underlying Python (e.g. Plotly) used to build the chart.
Data table Column headers + rows when the agent returns tabular results.
Key insights Text list; special markers may render as dividers between groups.
Suggested visualizations Cards showing title, chart type, dimension; click to request that chart from the agent.

Resizable divider — Drag between visualization and chat to give more space to either side.

Right: chat (Data Insights Agent)

The side chat is wired specifically to the Data Insights Agent (data_insights_agent). It supports:

  • Model selection from workspace-configured LLMs
  • Session id tracking so follow-ups stay in context
  • Context builder — each message can include current dashboard state (existing charts, datasets in play) so answers stay grounded
  • Streaming toggle and tool-grouping options where exposed

Events from chat to the dashboard (conceptually): new visualization, KPI, table, insights text, suggested charts; session changed / new session may clear or refresh the board so you do not mix unrelated analyses.

Core workflows

Goal Example prompt
Trend “Show me revenue by month for the last four quarters.”
Ranking “Top 10 products by units sold last month.”
Segmentation “Break down signups by channel and region.”
Quality “Highlight any weeks where conversion dropped more than 20% vs prior week.”
Follow-on After a chart: “Now split that by region” or “Export the SQL you used conceptually as documentation.”

Code transparency and reuse

Power users can open or copy the Python behind a chart to:

  • Re-run in a notebook
  • Adapt styling
  • Embed in internal reporting

This is intentional transparency: you see how the visualization was produced, not only the picture.

Model selection

Available models come from the same admin-configured pool as other agents (Model Settings). If responses are slow or low quality, try another model or narrow the question (smaller date range, fewer dimensions).

Troubleshooting

Symptom Likely cause What to try
Empty dashboard after prompt Agent still computing or returned text only Wait; ask explicitly for a chart or table
Chart fails to render Invalid spec or browser/WebGL limits Ask for a simpler chart type; retry
Stale panels after new session Session reset cleared state Expected—re-run the analysis
Wrong database or dataset Context not attached In other agents you attach DB context; here ensure the agent knows which connection/schema to use per your setup

Related

Last updated: February 21, 2026