Summary
Zach Blumenfeld, AI research engineer at Neo4j, leads a hands-on workshop at AI Engineer World’s Fair on using graph databases to give AI agents richer context in data lakehouse architectures. The session uses a fictional national auto repair chain — AutoFix Group — with data spanning BigQuery tables, cloud-stored vehicle manuals, safety bulletins, and repair logs, to illustrate why text-to-SQL and vector search alone fail when tables number in the hundreds or document stores reach millions of files.
The workshop demonstrates a graph data model that bridges the structured warehouse and unstructured document lake, enabling agents to answer both point queries (“how do I repair this specific vehicle?”) and estate-level questions (“what documentation do we lack for car types coming in this month?”) — the latter being nearly impossible to prove with semantic search alone. A key architectural decision is deterministic graph loading: rather than LLM-based entity extraction as used in Microsoft’s GraphRAG, Blumenfeld presents a containment tree structure reflecting document hierarchy — library → document → section — with ordering and cross-reference links, making loading idempotent, faster, and independent of LLM non-determinism.
Participants work through the full pipeline using Neo4j’s Graph Academy course, with credentials provided for BigQuery and the Anthropic API. The patterns demonstrated on BigQuery are shown to extend to Databricks and Snowflake environments. Ben Squire (senior developer advocate) and Ryan (partner architect) provide in-room support throughout.
📺 Source: AI Engineer · Published July 23, 2026
🏷️ Format: Tutorial Demo
