Model Manager¶

Availability: This agent is only available with Insights Hub offerings 4520 and 20420 (Standard Capability Package).

ModelManager is an AI assistant in Insights Hub for managing machine learning models end-to-end: discovery, registration, versioning, comparisons, deployment and inference.

User Journey Overview¶

The following images depict the user journey, describing how to prepare, register, deploy, and operationalize machine learning models with ModelManager.

Quick Start - Suggested Prompts¶

The following prompts illustrate common ModelManager tasks and recommended request patterns.

Capabilities¶

1. Discover & Inspect (Read)¶

The following table summarizes representative read operations supported by ModelManager.

Filters and pagination are supported on list actions:

List models with pageSize 10, sorted by name ascending

2. Register & Update (Write)¶

All create operations use a human-in-the-loop flow: 1. ModelManager collects your inputs (one missing field at a time). 2. It shows a JSON payload preview. 3. You confirm or request changes. 4. It submits only after your explicit "yes".

You can register resources either through a step-by-step workflow or by submitting the model, version, and artifact details in a single request.

Option A - Step-by-step registration¶

Step 1 - Register a Model

Register a model named Battery_coating_thickness, origin IDL,
description "Binary classifier for battery coating thickness"

Step 2 - Register a Version

Create version v2.0 for model Battery_coating_thickness,
description "Retrained on Q1 2026 data"

Step 3 - Register an Artifact

Note: Before registering an artifact, ensure the model file is physically present in the insights hub data lake at the exact storagePath you provide. ModelManager registers a reference to that path - it does not upload or validate the file's existence. Providing a path where no file exists will result in a non-functional artifact.

Create an artifact for version v2.0:
  name: "Battery coating model artifact"
  type: model-artifact
  storagePath: data/ten=demo/models/ClassificationModel/coating_thickness_classifier.bst
  format: xgboost
  cpu: 0.5, memory: 1Gi

Option B - Single-shot registration (model + version + artifact in one call)¶

Register model, version, and artifact in one shot:
  model name: Battery_coating_thickness, origin: IDL
  version: v2.0
  artifact name: "Battery coating model artifact", type: model-artifact,
  storagePath: data/ten=demo/models/RegressionModel/anomaly_detection.bst,
  format: xgboost, cpu: 100m, memory: 512Mi

Update an existing resource¶

Update description of model ID abc-123 to "Updated classifier"
Update artifact ID xyz-456, change memory to 2Gi

3. Custom Properties¶

Custom metadata can be attached to models, versions, and artifacts. Values are automatically wrapped in the correct format - you just provide plain key-value pairs:

Add custom properties:
  problem_type: Classification
  domain: Battery Manufacturing
  accuracy: 0.97
  optimized: true

Supported value types:¶

The following table defines the supported value types and their corresponding input examples.

4. Deploy a Model¶

ModelManager validates that a deployable artifact (model-artifact type) exists before proceeding.

Deploy model Battery_coating_thickness, version v2.0

Flow: 1. Resolves model and version IDs automatically. 2. Checks for a model-artifact - stops with a clear message if none exists. 3. Triggers deployment and reports the server ID and initial status. 4. Offers to check deployment progress.

5. Compare Versions¶

Compare version v1.0 and v2.0 of model Battery_coating_thickness

Returns a side-by-side comparison of the two versions' metadata and custom properties.

6. Run Inference¶

Run inference on server ID srv-abc with input:
  {"SprayPressure": 4.2, "Temperature": 22.5, "Humidity": 60}

If you don't know the server ID, ask:

List all model servers

Then use the server ID from the result.

Artifact Types Reference¶

The following table describes the supported artifact types and their intended purpose within a model version.

A version must have at least one model-artifact to be deployable.

Supported ML Frameworks and Artifact Formats¶

The Model Manager solution supports standard machine learning models exported in the following formats:

sklearn, tensorflow, pytorch, onnx, xgboost, lightgbm, keras, h5, pmml, pickle.

Models built with any of the frameworks above can be exported in their native file formats (for example, .bst for XGBoost, .joblib for scikit-learn) or serialized as pickle (.pkl), and then registered directly in the Model Manager using an agent.

Resource Sizing Guide¶

The following table provides representative CPU and memory values for deployment resource allocation.

If unsure, the default suggestion is cpu: 0.5, memory: 1Gi.

Error Reference¶

The following table lists common error conditions, their meaning, and the recommended corrective action.

Tips¶

• IDs are resolved automatically. You can use names instead of IDs in most prompts. ModelManager will look up the ID silently.
• Confirmation is required for all creates. Review the JSON preview carefully before confirming.
• Omit optional fields you don't need - don't send empty or null values.
• One question at a time. When inputs are missing, ModelManager asks for one field at a time to guide you through.
• Deployment requires a model-artifact. Register an artifact of type model-artifact before attempting to deploy.