The config layer

Config Studio

Point Config Studio at a database and it reads the schema and writes the config the agent answers from.

Config Studio is how a config gets written without anyone hand-authoring YAML. It inspects your tables, columns, and relationships, then drafts the whole thing: which tables matter, how they join, and what your team's vocabulary maps to.

Generate

Select a connection and start a generation. The run has three phases:

Profile

It introspects the schema (tables, columns, types, keys, and relationships), and samples values to understand what each column actually holds.

Adjudicate

It decides what belongs in the config: which tables are worth exposing, which joins are real, and which columns are the metrics people will ask for.

Emit

It writes the YAML and runs it through a verification gate. The gate is fail-closed: a config that does not verify is not offered to you.

Nothing is saved until you press Save. The draft is yours to read first, in plain language rather than raw YAML.

Refine

A config is not finished when it is generated. Refine improves an existing one from real evidence: the queries that came back wrong, and the thumbs-down feedback on answers.

Point it at a bad answer and it revises the config: a new version each time, so you can compare and roll back.

Custom tools

Beyond tables and metrics, Config Studio can carry tools the agent may call:

KindWhat it is
sql_macroA named, parameterised query the agent can invoke
pipelineA multi-step sequence over your data
pythonSandboxed Python for logic SQL cannot express

Tools follow the same rule as configs: nothing publishes unless it ran. You try it, see the real output, and then it is available.

Re-run after schema changes

The agent plans from the config, not from a live schema read. When you add or rename tables, re-run Config Studio: otherwise it is planning against a database that no longer exists in that shape.

On this page