AI Builder

The AI Builder (Claude-powered) lives in the pipeline builder create flow and as a floating widget on other app pages. Describe what you want to load — source, destination, optional dbt transform — and it generates a save-ready pipeline definition with inline credential fields where needed.

Who can use it

Only workspace roles with pipeline write (owners, solo users, and member invites) can save AI-generated pipelines. Viewers, catalog editors, and catalog browsers get a read-only message; they can still browse catalog docs and assets but not persist new pipelines.

What it generates

• Core pipeline — name, source type, destination type, connector-specific sourceConfiguration, and generated Python/YAML preview.
• Post-load dbt (EL+T) — when you ask for transforms after load, the AI sets sourceConfiguration.dbt with package_path, optional Git URL, schema, branch, and model selector — same shape as the form builder and canvas transform inspector.
• Linked dbt project — the AI can call list_dbt_projects and set dbtProjectId on the save payload to link a workspace project from /catalog/dbt (equivalent to the dbt project picker in builder/canvas).
• Python / SQL post-transform — optional post_transform block for scripts after load.

Human-in-the-loop save

1. Review the generated code preview in the chat.
2. Click Looks good to reveal save options.
3. Fill inline fields (e.g. GitHub org/repo) or skip and edit in the form builder.
4. After save, open the pipeline in the form builder or visual canvas.

Example prompts

• Load GitHub issues and PRs into Snowflake
• Sync Stripe payments to BigQuery with dbt staging
• GitHub → Snowflake EL+T with dbt models after load
• What workspace dbt projects do I have?

API

POST /api/elt/ai-assistant accepts { messages: [{ role, content }] } and returns message, optional savePayload (same as POST /api/elt/pipelines), requiredFields, and codePreview. Requires session auth and canWrite permission.

Pipelines · dbt transforms · Catalog