Skip to main content

Generate a validation with AI

Goal​

Describe what you want checked and get a ready-to-review validation: a name, a description, the assets it covers and the rules for each. You don't have to name the assets. Without them, the AI finds the relevant ones in your workspace.

Prereqs​

  • An AI engine. Native AI is in effect, or a BYOLLM provider is configured in Settings → AI. Generating spends the organisation's AI budget and counts against its hourly AI cap.
  • The editor role or higher in the workspace. Creating a validation and writing rules need can_manage_validation and can_edit_rule. A viewer sees Generate with AI disabled.
  • Profiles help but aren't required. With a profile, the AI grounds rules in the column statistics. Without one, it works from column names and types and stays conservative.

Steps​

  1. Open Validations and click Generate with AI.

  2. Describe what to check, in your own words. For example: "orders and refunds are complete, with valid amounts and ids".

  3. Pick the assets, or don't.

    • Pick one or more assets and the AI proposes rules for exactly those.
    • Leave the picker empty and the AI chooses the relevant assets from your workspace catalog. It sees names, tables, descriptions, tags and column names, never sample values.

  4. Click Generate. The draft is written in the background, usually within a few seconds. Nothing is saved yet.

  5. Review the draft.

    • Edit the name and description.
    • Remove any asset you don't want (× on its card). For an asset the AI picked, the card says why.
    • Under New rules, untick the ones you don't want and set each rule's severity.
    • Under Existing rules, tick the asset's current rules you want in the validation. Already on asset — reused marks a rule the AI proposed that the asset already has. It's included as the existing rule, not copied.
    • Notes at the top say what was left out and why. For example, invalid suggestions, or an asset restricted to another team.

  6. Click Create validation. This creates the validation and the new rules, and puts exactly the rules you kept in the validation. It all happens in one step: if anything fails, nothing is saved. The new validation opens.

  7. Run it from the validation page. Refine or remove rules that don't fit your data, since the AI only knew the statistics it was given.

Good to know​

  • Team restriction. If an asset is restricted to a team, the validation is restricted to that team too. Assets restricted to two different teams can't share a validation, so the draft leaves one out and says so.
  • Your access only. The draft is built with your workspace and team access. It never includes an asset you can't see, and only your team can see a draft that covers a restricted asset.
  • No double spend. Asking the same thing again within the hour returns the same draft.
  • Samples stay put. An asset that doesn't retain samples sends no sample values to the AI.

From an MCP client​

The MCP server has the same flow: mc_validation_generate starts a draft (assets optional), mc_suggestion_get polls it, and mc_validation_draft_accept creates the validation from the reviewed draft. The API endpoints are listed under Validation drafts.