How Recruit Holdings, the company behind Indeed, taught its text-to-SQL agent to read a website's hidden context
Summary
Recruit Holdings, the company behind Indeed, taught its text-to-SQL analytics agent to interpret button clicks by encoding website UX logic in YAML files that sit alongside dbt models. That gives the agent the hidden context of how a site works, so it can read user-behavior data the way the product team does.
Summary generated by brickster.ai. For the full article, follow the source link above.
More from dbt Labs Blog
How WMCHealth cut a 72-minute dbt pipeline down to 5 minutes, with one environment variable
WMCHealth cut a dbt pipeline from a 72-minute refresh to five minutes by switching one model between two execution modes with a single ENV_NRT environment variable: active records only as a table every five minutes, or the full dataset as an incremental overnight. The pattern runs from one codebase and cuts median runtime 93.6% (4,340 to 278 seconds) with over 99% less data processed per near real-time run, and it relies on inner-join propagation to carry the narrower scope through the DAG for free.
How WebstaurantStore used Snowflake Cortex AI to automate dbt staging for 700+ sources in 90 days
WebstaurantStore's three-person data engineering team onboarded more than 700 sources in 90 days, roughly a tenfold gain in ingestion speed by their own measure, by automating dbt staging from landing through the bronze layer with no added headcount. Snowflake Cortex LLM functions propose readable column names and business definitions from legacy OLTP metadata, and engineers review them in an intermediate table that the dbt Jinja macros read from instead of the raw AI output.
Meta:Context: a business context schema in dbt's semantic layer
Green Dot built Meta:Context, a five-layer business context schema inside dbt's semantic layer, so AI agents get more than a metric's value. The schema also carries what the number means, which gives agents the business context to interpret it correctly.
How PetScreening turned a dbt pipeline into a pet tag factory
PetScreening's FIDO Alert team connected an in-house tag-printing machine directly to their dbt project, treating it as another exposure fed by a gold-level model that also powers the growth team's dashboard, and now prints over 1,000 pet tags a day across two machines. Because bad data here means missed prints and reprints rather than a wrong dashboard number, a five-person team relies on dbt tests wired into Slack and a medallion architecture that let them swap data sources across three iterations without interrupting production.
How AlphaSense turned a 3-to-5-week data rebuild into a few hours with dbt and Euno's AI context
AlphaSense cut a data model rebuild from 3-to-5 weeks down to a few hours by pairing dbt governance with Euno's AI context platform. The same combination now supports their conversational analytics at scale.
Shrink the surface: governed self-service for analyst-owned dbt pipelines
Liftoff got self-service dbt to stick by shrinking what analysts are responsible for, rather than training them on more of the Databricks platform. The post lays out how that narrower scope makes governed, analyst-owned pipelines workable.