Skip to main content

The Insight Engine

The AI agent is your intelligent assistant for data analysis. Ask questions in plain English and watch as it builds complete analysis workflows automatically.

Standalone Conversations

For standalone data conversations outside of notebooks, see Conversations. Conversations support cross-database joins, data access controls, and sharing.

How It Works​

The Insight Engine appears in the right sidebar of any notebook. It can:

  • Write SQL queries from natural language descriptions
  • Create and execute cells to answer your questions
  • Build multi-step analyses with proper cell dependencies
  • Understand your data through schema introspection
  • Use semantic definitions for accurate, consistent metrics (coming soon)

Starting a Conversation​

  1. Open any notebook
  2. Click the chat icon in the right sidebar
  3. Type your question and press Enter

Example questions:

  • "Show me the top 10 customers by revenue"
  • "What's the month-over-month growth rate?"
  • "Create a chart showing sales by region"
  • "Join users with orders and find the average order value"

Tool Approvals​

For sensitive operations, the agent asks for your approval before executing. You'll see a summary of what it wants to do:

  • Creating cells — Shows the SQL or content to be added
  • Executing queries — Shows which cells will run
  • Modifying structure — Shows proposed changes
  • Running SQL (in conversations) — Shows the query before execution
  • Derived queries (in conversations) — Shows the cross-database join query

Click Approve to proceed or Reject to cancel. Tool approvals apply in both notebooks and Conversations.

Derived Queries​

In Conversations, the agent can combine results from different databases using derived queries:

  1. The agent runs queries against individual databases
  2. Results are loaded into DuckDB in-memory
  3. A derived query joins, filters, or aggregates across those result sets

This enables cross-database analysis without ETL pipelines. Derived queries are composable — they can reference other derived queries.

Conversation Management​

You can have multiple conversations per notebook. Each maintains its own context and history. For standalone conversations outside of notebooks, see Conversations.

Grounding with Semantic Layers (Coming Soon)​

When the Semantic Catalog launches, the agent will become more accurate with your defined semantic layers. These will provide:

  • Metric definitions — How to calculate revenue, churn, etc.
  • Dimension mappings — How to group and filter data
  • Join paths — How tables relate to each other
  • Business context — What terms mean in your domain

Best Practices​

Be Specific​

Instead of "show me sales", try "show monthly sales revenue for 2024, grouped by product category"

Iterate​

Start with simple questions, then ask follow-ups to drill deeper

Use Your Schema​

Reference actual table and column names when you know them

Define Metrics (Coming Soon)​

Create semantic layers for metrics you use frequently — see Semantic Catalog

Agent Studio​

Monitor your AI usage in Agent Studio:

  • Total conversations and token usage
  • Usage by team member
  • Cache efficiency (how often cached responses are reused)
  • Conversation history and search

This helps optimize costs and understand how your team uses the AI.

Limitations​

  • The agent works best with well-structured schemas
  • Very complex queries may need manual refinement
  • Large result sets are sampled for display
  • Some operations require tool approval