Model Vibe and Insights

datahashi has two conversational AI products. Model Vibe authors models in Studio. Insights answers business questions on published models.

Both agents use structured tools. A human stays in the loop. AI never writes SQL to your warehouse. AI never approves a model. AI never publishes a model.

flowchart TD
    subgraph MV["Model Vibe (authoring) · conversation model_create"]
        direction TB
        M1["Introspects table schema<br/>Drafts entity and metric YAML<br/>Generates reviewable Proposals (diffs)"]
    end
    subgraph IN["Insights (analytics) · conversation insights"]
        direction TB
        I1["Discovers the catalog<br/>Runs governed queries<br/>Suggests charts, explains trends"]
    end
    M1 --> A["Human editor approval<br/>review the diff, then publish"]
    I1 --> V["Verified figures only<br/>no invented SQL"]

The two products have different jobs:

• Model Vibe helps you design and edit YAML before publish.
• Insights helps you query and chart published metrics.

Model Vibe

Model Vibe sits in Studio next to the YAML editor.

How Model Vibe works

1. Read the schema. Vibe calls Query Runner POST /v1/schema. It reads tables, primary keys, and foreign keys.
2. Design in conversation. You ask in plain English. Example: “Create metrics for gross margin and monthly active subscribers.”
3. Emit a Proposal. Vibe does not overwrite your draft. It emits a typed Proposal with file diffs.
4. Review and apply. An editor inspects the diff and clicks Approve. The changes merge into the draft.
5. Publish. An authorized editor publishes the draft to the engine.

Safety constraints

• AI cannot approve proposals.
• AI cannot publish models.
• Proposal changes run through model validation before they merge.

Insights

Insights is a natural-language interface over published data.

How Insights works

1. Ground in the catalog. Insights reads the caller’s granted metrics and dimensions. It sees default time dimensions and entity relationships.
2. Emit a Semantic Query. Insights does not write raw SQL. It compiles the question into metrics, dimensions, and filters.
3. Run and combine. For cohorts or multi-step questions, Insights runs several governed queries. It then explains the difference.
4. Return structured blocks:
   • text: Narrative around the numbers.
   • stat: Key metric highlights.
   • query_result: Table with suggested charts (bar, line, area).
   • callout: Anomalies or extra context.

Why Insights does not invent numbers

Insights sends Semantic Queries to the engine. Every query gets:

• Fan-out-safe compiler joins.
• Server-side data roles and row-level security.
• DuckDB rollup acceleration.
• An immutable audit log on Query Runner.

Next: PostgreSQL wire protocol · Write your first model.