AI/BI Accelerator¶

The AI/BI accelerator scaffolds a complete Databricks AI/BI solution over the TPCH sample dataset: a metric view as the semantic layer, a Lakeview dashboard for visual analytics, and a Genie Space for natural-language querying — all deployed via a single Databricks Asset Bundle.

What gets generated¶

ai-bi/
├── databricks.yml                                        # Asset Bundle root config
├── .gitignore
├── resources/
│   ├── dashboards/
│   │   ├── dashboard.yml                                 # Lakeview dashboard resource
│   │   └── tpch_overview.lvdash.json                     # Dashboard definition (3 pages)
│   ├── genie_spaces/
│   │   └── tpch_genie.genie_space.yml                    # Genie Space resource + config
│   ├── jobs/
│   │   └── setup_views_job.yml                           # One-time job to deploy metric view
│   └── warehouses/
│       └── warehouse.yml                                 # Serverless SQL warehouse
└── notebooks/
    └── setup_metric_views.py                             # Creates the metric view at runtime

Metric view — `v_tpch`¶

A single metric view sourced from samples.tpch.orders, joined to customer and nation:

orders → customer → nation

Filter: o_orderdate >= '1995-01-01'

The metric view acts as the semantic layer for both the dashboard and the Genie Space. Fields carry display_name and synonyms metadata so Genie can resolve natural-language terms like "revenue", "sales", or "AOV" to the correct measure without additional configuration.

Fields (dimensions)

Field	Expression	Description
`order_date`	`o_orderdate`	Date the order was placed
`order_month`	`DATE_TRUNC('MONTH', order_date)`	Month grain for trend analysis
`order_year`	`YEAR(order_date)`	Year grain
`order_status`	CASE on `o_orderstatus`	Open / Processing / Fulfilled
`order_priority`	`SPLIT(o_orderpriority, '-')[0]`	Priority number (1–5)
`customer_name`	`customer.c_name`	Customer display name
`market_segment`	`customer.c_mktsegment`	AUTOMOBILE / BUILDING / FURNITURE / HOUSEHOLD / MACHINERY
`customer_nation`	`customer.nation.n_name`	Customer country (synonyms: nation, country)

Measures

Measure	Expression	Format
`order_count`	`COUNT(DISTINCT o_orderkey)`	compact number
`total_revenue`	`SUM(o_totalprice)`	compact USD (synonyms: revenue, sales)
`unique_customers`	`COUNT(DISTINCT o_custkey)`	compact number
`avg_order_value`	`MEASURE(total_revenue) / MEASURE(order_count)`	USD (synonym: AOV)
`revenue_per_customer`	`MEASURE(total_revenue) / MEASURE(unique_customers)`	USD
`open_order_revenue`	`SUM(o_totalprice) FILTER (WHERE o_orderstatus = 'O')`	USD (synonym: backlog)
`fulfilled_order_revenue`	`SUM(o_totalprice) FILTER (WHERE o_orderstatus = 'F')`	USD
`t7d_customers`	`COUNT(DISTINCT o_custkey)` trailing 7-day window	compact number

Lakeview dashboard¶

Three pages backed by MEASURE() queries against v_tpch:

Overview — Total Revenue, Total Orders, Avg Order Value counters; monthly revenue trend (area chart); revenue by market segment (bar) and by priority (pie)
Order Analysis — Revenue and count by order status (bar + pie); revenue by country (table); unique customers counter
Top Customers — Top 20 customers by total revenue (table with segment, country, avg order value); revenue by segment for those customers (bar)

Genie Space¶

Pre-seeded with 10 curated sample questions and domain-specific instructions tuned to the TPCH dataset. Genie uses the metric view's display_name and synonyms fields for semantic resolution, so users can phrase questions naturally:

"What is the total revenue?"
"Which market segment generates the most revenue?"
"What is the revenue split between open, processing, and fulfilled orders?"

Why the setup job is needed¶

DAB does not yet support metric_views as a native resource type. The setup job runs notebooks/setup_metric_views.py on serverless compute, which executes a CREATE OR REPLACE METRIC VIEW statement dynamically using the catalog and schema variables. This means catalog and schema can be overridden at deploy time without regenerating the project:

databricks bundle deploy --target prod --var catalog=prod_catalog --var schema=tpch_metrics

The dashboard and Genie Space depend on v_tpch existing — run the setup job once before opening either.

Note

The Genie Space's description and example-question SQL also use ${var.catalog}/${var.schema}, so they follow the same deploy-time override. The Lakeview dashboard (tpch_overview.lvdash.json) is the one exception: dashboard content is an opaque JSON blob that DAB does not variable-substitute, so its queries are baked to the catalog/schema config values at dpa init time. If you need a different catalog for the dashboard, edit the JSON's query fields directly after scaffolding.

Requirements¶

Databricks CLI v1.3.0+ (required for Genie Space bundle support)
Unity Catalog enabled in your workspace
Access to samples.tpch (built-in Databricks sample data — no setup needed)

Usage¶

dpa init ai-bi
cd ai-bi

# Deploy resources (warehouse, dashboard, genie space)
databricks bundle deploy

# Create the metric view (run once before opening the dashboard or Genie Space)
databricks bundle run ai_bi_setup_views

For production:

databricks bundle deploy --target prod
databricks bundle run ai_bi_setup_views --target prod

Variables¶

Variable	Default	Description
`catalog`	`dpa_ai_bi_dev`	Unity Catalog catalog for the metric view
`schema`	`tpch_metrics`	Schema for the metric view
`warehouse_id`	(required)	SQL Warehouse ID for the setup job and Genie Space

Set warehouse_id in databricks.yml before deploying (find it under SQL Warehouses → your warehouse → Connection details). Override other variables at deploy time:

databricks bundle deploy --var="catalog=my_catalog" --var="schema=my_schema"

Customising the Genie Space¶

After deploying, you can refine the Genie Space in the Databricks UI — add example SQL queries, adjust instructions, or add more sample questions. To pull changes back into the bundle:

databricks bundle generate genie-space --resource ai_bi_genie --force

Warning

Redeploying overwrites any existing Genie Space with the same name, including chat history. Use dev/prod targets to keep environments isolated.