AI on Your Lakehouse: Context Comes in Shapes, Not Queries
Build a graph knowledge layer over your documents and warehouse data so your agent can use them reliably
In this 1-hour course, you will learn
Why your agent can struggle to use your data reliably - and how graph shapes can fix it.
Vector search and Text2SQL each hand your agent a slice of data - similar passages, or a guessed SQL join - but not the view to know what is truly relevant and how it connects. Answers come back confident but wrong. The problem isn't caused by a bad model or bad query, but rather a lack of context, and context comes in shapes.
In this hands-on workshop, you give an agent three reusable shapes:
- Connections (Paths) - how structured warehouse data joins and connects
- Table of Contents (Trees & Links) - navigate documents
- Themes (Communities) - surface patterns nobody named
These shapes help answer the questions many modern stacks struggle with and get quietly wrong - not just the single lookups, but also the estate-level ones: the patterns across an entire document set, what your records show happening that the documentation never covers, and how things connect across a boundary no one query spans. Each needs the agent to cover, navigate, or follow the whole structure - not retrieve a similar slice.
You work the pattern on AutoFix Group, a fictional auto-repair chain: manuals, bulletins, and recalls as PDFs alongside vehicles, work orders, and parts in a warehouse. In a hosted Codespace, you and your coding agent build a service-advisor skill shape by shape, then put it to work across both halves - federating the warehouse with the graph instead of migrating it.
The pattern is portable by design: the warehouse here is BigQuery, but swap the connector and the same shapes work on Snowflake, Databricks, or anywhere your data lives.
Before taking this workshop, you should have:
- A basic understanding of AI agents
- The ability to read and run basic SQL queries
- Familiarity with data warehouse or lakehouse concepts (tables, keys)
For the hands-on path you will need a coding agent (this workshop assumes Claude Code) together with the Codespace - which provides read-only access to the BigQuery dataset and your Neo4j instance. The embedded Query pane lets you inspect and query the graph by hand as you go.
Why agent context is a problem of shapes, not queries - and where vector search and Text2SQL often fall short
How a graph semantic layer (connections shape) improves Text2SQL
How to navigate documents with trees and links
How to surface themes with Leiden community detection
How to get an agent to answer estate-level questions - covering whole document libraries, identifying what is undocumented, and connecting documents to records across the structured/unstructured data boundary
How to write graph queries via agentic coding - using neo4j-cypher-skill with a spec instead of hand-writing Cypher
How to use the Neo4j CLI for graph reasoning, and make it available to coding agents