How Genie works

Introduction to Databricks Genie

Gang Wang

Senior Data Scientist

Genie as a compound AI system

nanobanana: half: Three team members collaborating - a Brain character representing the LLM, a Map character representing Unity Catalog, and a Muscle character representing the SQL Warehouse - working together as a unified system

A team working together:

The Brain (LLM) - Understands your language
The Map (Unity Catalog) - Knows where the data lives
The Muscle (SQL Warehouse) - Does the heavy lifting

Unity Catalog integration

hierarchy: Unity Catalog, Schema, Tables, Columns

What your data team provides:

Table descriptions - What data is stored
Column comments - What each field means
Relationships - How tables connect

How Genie processes your question

Pipeline step 1: Analyze

Step 1: Analyze your natural language to understand your intent - not just keywords, but what you actually want to know.

How Genie processes your question

Pipeline step 2: Analyze and Lookup

Step 2: Look up relevant tables and columns from Unity Catalog, finding the data sources that can answer your question.

How Genie processes your question

Pipeline step 3: Analyze, Lookup, and Generate

Step 3: Generate SQL with the right joins, filters, and aggregations.

How Genie processes your question

Pipeline step 4: Full pipeline

Step 4: Execute that SQL against your data warehouse and format results into tables and visualizations.

Question → Answer in ~2 seconds

When Genie asks for clarification

nanobanana: full: A conversational interface showing a user asking an ambiguous question and the AI assistant responding with a clarification prompt offering multiple interpretation options to choose from

Genie is "accurate over assertive"

It asks when facing a business choice, not a grammar fix
Example: you ask for "Sales" and Genie sees both Gross_Sales and Net_Sales - it won't guess

This is a feature, not a limitation!

Demo: Genie clarification prompt in Bakehouse

Let's practice!

Introduction to Databricks Genie