4. Generate a model

After every successful scan, Keel generates the semantic model for each model that includes the source (automatic generation is on by default). A workspace with no model gets one — "Semantic model", authored by Keel — on its first scan. You can also run it on demand: POST …/models/{id}:generate.

What generation uses

• Accepted proposals, plus High-confidence undecided ones by default (joins ≥ 0.85, table classifications ≥ 0.6). Rejected proposals never.
• Staging and unclassified tables are left out, with a note saying why.

What it produces

• Entities for facts, dimensions and slowly changing dimensions, with dimensions (and time grains), measures, hierarchies (from functional dependencies proven on the profile sample) and segments.
• Joins with cardinality and roles — including role-playing dates such as order, due and ship date.
• Metrics: count measures and simple metrics; with an LLM, suggested ratio and derived metrics (names only, never values) stored as Draft with origin LlmSuggested, confidence ≤ 0.7 and the rationale.
• Every object carries provenance (origin, confidence, rationale), so generated objects stay identifiable.

On AdventureWorksDW, generation produces 29 entities, 47 joins and 36 metrics with no validation errors.

Where it goes

Generation never touches main. It commits to a branch named keel/proposals-{run} and opens (or re-checks) one change request. Objects you defined yourself on that branch win over regenerated ones. Re-running generation for the same scan is idempotent: it commits only when the content changed.

Look at it

• Change requests › #n — the object-level diff, the checks, reviewers.
• Workspace (/models/:id/workspace) — the model as a graph: entity cards with typed fields, crow's-foot joins, domain groups, trace path, problems panel, a YAML code view and SQL preview. Edits autosave to your private working copy and commit to a branch (main is protected). Proposed objects are drawn dashed with their confidence.
• Strategy (/models/:id/strategy) — see below.

The strategy report

Keel recommends how to deliver the layer, not only what is in it. It measures the schema's shape (star-ness, normalization depth, subject areas, volumes, cross-source joins) and scores four patterns 0–100 with a rationale: Star-schema centric, Entity / metrics-first, Virtual / federated and One big table. It also suggests where to serve (from a short questionnaire), derived facts for header/detail pairs (with the SQL, in the source's dialect), readiness, and up to ten next actions you can tick off.

Next: 5. Release it.