Sign in Start free
Ship a new version

Ship a new version

Ship a retrained model as a new version, compare what the two learned, and retire the old one.

A version is a copy of a model under the same lineage_id: same schema, outputs, inference, reference data and structure, with its own slug, its own history and its own insights. The version that produced the old predictions and explanations keeps them.

  1. Create the version with POST /v1/models/{model}/versions, or upload the retrained estimator with version="2" and a new slug.
  2. Replace what changed: PUT …/structure for new weights, PATCH for a schema change, PUT …/reference-data for a new validation set.
  3. Run full analyses on both and compare their insights; GET /v1/models/{model}/versions lists the lineage, and the project overview's feature matrix shows which inputs each version leans on.
  4. Point traffic at the new slug. Delete a version with DELETE …/versions/{version} when it is retired.
v2 = client.models.create_version("credit-risk-v3", version="4", description="Adds utilisation trend features.")
client.models.upload_structure(v2.id, clf_v4)                      # the retrained estimator
client.models.upload_reference_data(v2.id, rows=X_valid_v4.to_dict("records"), labels=y_valid_v4.tolist())
client.models.update(v2.id, slug="credit-risk-v4", features=FEATURES_V4)

for ref in ("credit-risk-v3", "credit-risk-v4"):
    client.analyses.wait(client.analyses.create(model=ref, type="full").id)
    importance = client.insights.list(model=ref, type="global_importance").data[0]
    print(ref, [(r["feature"], r["share"]) for r in importance.payload["features"][:3]])

for version in client.models.versions("credit-risk-v3").data:
    print(version.version, version.slug, version.explanations_count)
curl https://api.kaimei.ai/v1/models/credit-risk-v3/versions \
  -H "Authorization: Bearer $KAIMEI_API_KEY" -H "Content-Type: application/json" \
  -d '{"version": "4", "slug": "credit-risk-v4", "description": "Adds utilisation trend features."}'