This guideline will help you to activate and use Model Catalogue. Model Catalogue is a company-specific catalogue for the models that are developed and finetuned in AI/ML Development and MLOps platforms. Currently Saidot supports integration and governance of models developed in Azure AI Foundry, Microsoft Azure Machine Learning, Azure AI Services and Azure OpenAI Services, AWS Bedrock and Google Gemini Enterprise Agent Platform. .
The Model Catalogue offers Data Scientists and AI developer possibility to govern their ML & AI models at Saidot by
-
Integrate to ML & AI development or MLOps platform to govern models at Saidot
-
Import models through native integrations, REST API or MCP Governance connector
-
Govern models by adding information including description of performance, limitations, allowed and not allowed use and mark model approvals
-
Manage links to systems, datasets and model specific risks to inherit datasets and risks automatically to systems
-
Configure event-based governance actions for models
Decide how to configure Model Catalogue
Saidot Model Catalogue can be populated through different routes:
-
Native integrations to AI/ML-development platforms to import actual model deployments as specific model cards (Saidot pulls the metadata from the platform).
-
REST API connection to import model cards through integration layer that pushes the model metadata to Saidot Model Catalogue.
-
Saidot Governance MCP connection to create platform agnostic model cards using specific asset registration skills and prompt templates. These model cards could be linked across multiple Platform-level systems. You can learn more about integrated and agentic workflows here.
Model Governance can be configured in multiple ways:
-
Add custom fields to Model cards for model risk tiering and collecting needed governance data
-
Configure event-driven governance actions for monitoring and observability
Import Models through native integrations
Saidot Azure integrations are enabled by Admins using Admin interface before they can be enabled in Integrations. Other integrations are activated directly in Integrations. Please visit the Integration guideline for further information on activating your Model and Agent catalogues.
If your organisation Saidot Admins and Space Managers have not created integrations for Model catalogue, it will be empty.
The integration view contains information about available in-built integrations to your ML/AI development platforms. In addition to these integrations, Saidot allows to integrate the model data also through an API as an integrated workflow. You can learn more about the integrated workflows here.
After successful integration, models from Azure OpenAI Service, Azure AI Services and Azure Machine Learning can be imported and governed at Saidot.
Step 1. Import models
Start the model import by selecting Models and Import Models. Wait for Saidot to get the models to be imported from the platforms with established integrations. This might take a couple of minutes.
Select the models to be imported to the Model Catalogue.
Imported models will be visible in the Model Catalogue. You can search and filter models based on their metadata, including:
-
Owner
-
Integration (the model card source)
-
Model type (base model or custom model)
Step 2. View Model Card
View the deployment and model details in the Model Card. This information is integrated from the development platforms directly. The fields are different based on the platform-specific metadata on the specific model and deployment. When model is imported, it does not yet contain links to agents, systems or datasets. Risks are linked automatically to the model card in case the Base model exists in Saidot Library.
Step 3. Link systems, agents and datasets
Model card allows you to manage links to agents, systems and datasets. The links to datasets and agents ensure that they are being added to System when the model is linked to it’s Components.
The links to Systems can also be managed through the Model card.
Step 4. Govern model
Oversee and manage your model on the Model Governance section including
-
Model owner
-
Model risk level that can be inherited to AI system using automated workflow
-
Company-specific Custom fields that can be configured to model cards
-
Approval status
Model developers can also record specific information about the accepted use policy and upload or link evidence.
-
Capabilities (the intended use of the model)
-
Performance and limitations
-
Accepted and unaccepted use
-
Evidence documents
Step 5. Link risks
If the model is built on an existing model in the Saidot Model Library, model specific provider reported risks will be automatically connected to the model in the Model Catalogue. Model owner can add additional risks, for example related to model performance and use. These risks will be linked automatically to all systems using the model.
Step 6. Link model to System
Link the model to the AI system
When linking the mode, you are asked to describe the tasks. This will enable you to generate relevant evaluation plans for the model deployment. Currently automatic evaluation plans age generated only to models imported through Azure integrations.
Datasets will be automatically linked from model to the system or they can be added manually from the Dataset catalogue.
Step 7. Manage inherited model risks in system context
Evaluate the model risks in System context by setting inherent risk level, treatment strategy, linking controls, evidence and evaluating the residual risk level. Learn more about risk management here.
Model cards created through Governance MCP
Model cards can be created also without integrations through Governance MCP. You can use specific skills or prompt templates to create standard model cards, based on your internal guidelines. Typically model cards created through MCP are platform-agnostic and used in situations where they need to be linked as approved models across different development platforms. When you generate a model card through MCP, it does not automatically populate the risks from Library. You can specify in the prompt how you want to use the Model Library risks in these model cards. The MCP created model cards will be labelled as Custom models.
Model cards created through REST API
Model metadata can be also pushed through the selected integration layer via Saidot REST API to Model Catalogue. This option is optimal in cases where there are integration development reources available, preference to use “push” instead of “pull” and integration method and automate the model card creation workflow without manual import action. These models will be also labelled as custom models.
Model monitoring and observability
The trigger based governance workflow allows you to connect observability events to models.
This workflows enables you to track model related changes and specify the targeted notifications.