Jaffle Logistics

Jaffle Logistics is a fictional logistics company, and its data looks like most companies' data: scattered.

One account's story lives across roughly eight disconnected systems: a CRM, a support ticketing tool, dispatch notes from the ops floor, incident reports, legal contracts, Slack threads, and call transcripts. Some of those systems reference a shipment by a clean, structured ID. Others bury it in a sentence of free text. Some don't reference a shipment at all; they reference an incident that itself references a shipment.

Ask a targeted question, "what contributed to poor delivery performance for the Jaffle Equipment account this year," and the honest answer is spread across every one of those systems, in a mix of structured columns and prose nobody indexed for this purpose.

This project is a worked example of a real ceiling in exact-match SQL patterns, joins and regex, and what it actually takes to get past that ceiling while staying in dbt:

• The problem shows the deterministic baseline: how far you can get with joins, a shipment ID regex, and a lot of patience, and where that approach runs out of road.
• Context engineering walks through dbt_context_engineering, a package that chunks, embeds, and enriches the same free text in SQL, orchestrated by dbt, so it becomes searchable by meaning instead of exact match. Enrichment runs alongside embedding, not after it. It's a real building block, still short of a finished RAG pattern.
• Comparison puts both approaches side by side against real output, including where plain vector search falls short.
• Governance covers what keeps an AI-function layer like this safe to run: cost guards, an audit log, and an incremental design that means reruns cost nothing once the content stops changing.
• Multi-platform documents what it took to get the same pipeline running correctly on Snowflake, BigQuery, Databricks, and DuckDB.
• Roadmap is honest about what's deliberately not built yet.