AI on Your Lakehouse: Context Comes in Shapes, Not Queries
Neo4j lakehouse workshop: build a graph knowledge layer over PDF documents and BigQuery data so your AI agent answers estate-level questions reliably.
In this 1-hour workshop, you will learn
AI on Your Lakehouse is a hands-on Neo4j workshop about why agents struggle to use enterprise data reliably - and how graph shapes fix it. Vector searchFinding the records whose vectors lie closest to a query vector. and Text2SQL each hand your agent a slice of data: similar passages, or a guessed SQL join. The problem is not a bad model or a bad query - it is missing context, and context comes in shapes.
In this workshop you give an agent three reusable shapes: connections (paths) showing how structured warehouse data joins, a table of contents (trees and links) for navigating documents instead of just searching them, and themes (communities) that surface patterns nobody named. The workshop is written for instructor-led sessions, but everything runs in a hosted Codespace, so you can also work at your own pace.
Agent Context Is a Problem of Shapes, Not Queries
Understand why vector search and Text2SQL fail to capture relational context, and learn how graph shapes give agents the structure they need.
Graph Semantic Layer for Text2SQL
Build a graph semantic layer that exposes table relationships as connections, enabling agents to generate accurate SQL across complex schemas.
Document Navigation with Trees and Links
Model PDF document structure as a graph of trees and cross-links so agents can traverse chapters, sections, and references without losing context.
Theme Discovery with Leiden Community Detection
Apply Leiden community detection to your document graph to surface hidden themes and cluster related content across large unstructured libraries.
Estate-Level Question Answering
Configure an agent to answer questions spanning entire document libraries, flag undocumented gaps, and join unstructured documents to structured records.
Graph Queries via Agentic Coding
Write Cypher using the neo4j-cypher-skill spec and an agentic coding workflow instead of authoring queries by hand, reducing errors and iteration time.
Neo4j CLI for Graph Reasoning
Use the Neo4j CLI to run graph reasoning tasks and expose that capability to coding agents so they can query and traverse the graph autonomously.
What you need to take part.
A GraphAcademy account
The workshop is delivered here on GraphAcademy, so you need to be signed in to work through the lessons and keep your progress. Creating an account is free.
A GitHub account, or Git on your own machine
You write code against the genai-workshop-lakehouse repository. The quickest route is a GitHub Codespace, an online editor that clones the code and installs everything for you — that needs a GitHub account.
If you would rather work locally, clone the repository with Git and run it in your own editor instead. No GitHub account is needed for that.
7 modules, 1 hour.
Who this workshop is for
This workshop is for AI engineers, data engineers, and architects who have built retrieval over their data and hit its limits - the agent reaches the data, but you cannot trust its answers. You direct a coding agent throughout (the workshop assumes Claude Code) while you make the decisions about shape, scope, and correctness. You should understand AI agents at a basic level, read and run simple SQL, and know warehouse concepts such as tables and keys. You do not need to hand-write CypherNeo4j's implementation of GQL, the ISO standard query language for graph databases. It is declarative: you describe the pattern to find, and the database decides how to find it.; if graphs are entirely new, Neo4j Fundamentals is an optional primer. The Codespace provides read-only access to the BigQuery dataset and your own Neo4j instance, with nothing to install locally.
What you'll do
You work the pattern on AutoFix Group, a fictional auto-repair chain on a cloud lakehouse: repair manuals, service bulletins, and recall notices live as PDFs in cloud storage, while vehicles, work orders, parts, and diagnostic trouble codes live in BigQuery tables. Shape by shape, you and your coding agent build a service-advisor skill over that estate. You start on a single car in the bay - "what fixed this code on cars like this one?" - then zoom out to the estate questions: undocumented failures in the field, patterns across every bulletin and recall, and how a fix traces from documents to work orders.
Each shape carries a key decision: where it should live. The connections graph is a semantic layer built from the warehouse's metadata only - table names and foreign keys - so the rows stay in BigQuery while the documents become graphs in Neo4j. You federate rather than migrate, and the pattern is portable: swap the connector and the same shapes work on Snowflake, Databricks, or anywhere your data lives.
Where to go next
If GraphRAGRetrieval-augmented generation whose context comes from a knowledge graph, so the model can follow the relationships between facts. is new to you, Neo4j & GenerativeAI Fundamentals covers GraphRAG from the ground up. Building Knowledge Graphs with LLMs goes deeper into turning documents into graphs, Get started with Graph Data Science covers the community detectionA family of algorithms that group nodes by how they connect. Each algorithm defines a community differently. behind the themes shape, and Developing with Neo4j MCP Tools shows you how to connect agents to Neo4j. Find upcoming instructor-led sessions and the other companion courses on the workshops page.