Beyond standard metrics: extending dbt semantic models for high-accuracy Text-to-SQL
Breakout sessionPowering AI-ready dataExpertData leadersRetailMedia
Despite the GenAI boom, Text-to-SQL still requires manual query review, blocking true automation. The root cause is context: raw data schemas overwhelm LLMs, while basic prompts lack necessary detail.
dbt Semantic Models offer a great foundation for context. However, a technical gap exists: their standard API is built for BI tools, outputting data in formats that LLMs cannot easily digest.
To solve this, we custom-extended the dbt API into a truly "AI-Friendly API." Our system translates semantic definitions into a natively LLM-consumable format. Within this extension, we also integrated crucial business metadata—like cardinality and domain rules—giving AI precise context without fragile prompt engineering.
To prevent context decay, we automated synchronization between our dbt models and this extended semantic layer.
This session explores the architecture behind our AI-Friendly API and how it dramatically improved our Text-to-SQL accuracy and workflow efficiency.
dbt Semantic Models offer a great foundation for context. However, a technical gap exists: their standard API is built for BI tools, outputting data in formats that LLMs cannot easily digest.
To solve this, we custom-extended the dbt API into a truly "AI-Friendly API." Our system translates semantic definitions into a natively LLM-consumable format. Within this extension, we also integrated crucial business metadata—like cardinality and domain rules—giving AI precise context without fragile prompt engineering.
To prevent context decay, we automated synchronization between our dbt models and this extended semantic layer.
This session explores the architecture behind our AI-Friendly API and how it dramatically improved our Text-to-SQL accuracy and workflow efficiency.
