Stop your agent guessing what status means
An LLM writing SQL against your warehouse has three bad options: guess from the column name, read the SQL, or ask a human. Give it a fourth.
Start with this recipe →Just manifest.json. No warehouse credential, no dbt Cloud token, no writes to your repo.
Opens the CoreModels app at this recipe. Sign in with Google or Microsoft.
For agents, no sign-in: /recipe/agent-grounding.json · /recipe/agent-grounding.md
Every text-to-SQL tool and coding agent pointed at your warehouse is inferring meaning from column names and whatever free text happens to be in a description. It cannot know that status has exactly five permitted values, that return_pending is a kind of return rather than a peer of it, that customer_id means the same thing here as cust_id does two models over, or that this column is the one bound to a public term the rest of the industry uses. So it guesses confidently, and a confident wrong answer about meaning is worse than no answer. This workspace makes the governed definitions readable by machines: an agent connects over MCP and can ask what a column means, which values are permitted and what each one is, what the column is bound to externally, and what else carries the same concept — and it gets the governed answer a human curated, not an inference. dbt's own agent surfaces expose dbt's contents; this exposes the meaning behind them.
What you do
- Import the manifest; each accepted_values list becomes a governed vocabulary.
- Fill three fields (meaningNote, commonMistake, doNotUseFor) on the columns agents get wrong. About 20 is enough to test the loop.
- Connect your assistant to the read-only MCP endpoint (it reads your governed definitions, not your warehouse) and replay a question it used to get wrong.
Guides for this recipe
Strategy
Why the problem exists, and what changes once it is fixed.
Engineering
How it works: the import, the API, the MCP tools, the CI loop.