You can integrate AI models and agents from different data sources in Data Catalog. These integrations are only available via Edge.
You can configure Edge connections and capabilities without an active AI Governance license. However, AI Governance must be enabled to harvest AI model and agent metadata, ingest corresponding AI assets in Data Catalog, and access the dashboards and features necessary to visualize and govern your AI landscape.
AI integrations
Collibra offers out-of-the-box integrations that ingest machine learning (ML) model and AI agent metadata from the following AI platforms:
- Anthropic AI
- AWS Bedrock AI
- AWS SageMaker AI
- Azure AI Foundry
- Azure ML
- Databricks AI
- Gemini Enterprise Agent Platform
- MLflow AI
- OpenAI
- SAP AI Core
- Snowflake Cortex AI
Traceability and training data access
What Collibra can trace back to an AI model or agent differs by integration. The integrations link the relations that the source platform reports; it doesn't compute lineage itself.
- Input data attaches to the AI Model Version asset.
- Output data attaches to the AI Model Deployment asset.
- Some integrations separately link the AI Agent to its training data.
| Integration | AI Model Version (input data) | AI Model Deployment (output data) | AI Agent (training data link) |
|---|---|---|---|
| Anthropic AI | Not supported | Not supported | Not supported |
| AWS Bedrock AI | Yes, from an S3 bucket represented as a Storage Container or File asset | Yes, to a Storage Container or File asset representing the S3 location where inference results land | Not supported, knowledge bases become AI Agent Tool assets via the AgentVersionCallsTool relation |
| AWS SageMaker AI | Not supported | Yes, to a Storage Container asset representing the S3 folder or object where inference results are written | Not supported, no AI Agent asset type exists in this integration |
| Azure AI Foundry | Yes, from a Storage Container asset representing the training dataset | Not supported, the AI Model Deployment asset exists but does not have an output data relation | Yes, automatic, the AI Agent links directly to a Storage Container asset it uses |
| Azure ML | Not supported | Not supported | Not supported, no AI Agent asset type exists in this integration |
| Databricks AI | Yes, from a Delta Lake table | Yes, to a Delta Lake table that stores the output | Yes, automatic, the AI Agent Tool asset links to the table or volume it reads from |
| Gemini Enterprise Agent Platform | Yes, from a Google Cloud Storage file or BigQuery table | Yes, to a BigQuery table holding the output | Not supported, the AI Agent's tool usage isn't linked to any file or table asset |
| MLflow AI | Yes, only for Delta table-sourced training data | Not supported, no deployment asset type exists in this integration | Not supported |
| OpenAI | Not supported | Not supported | Not supported, no AI Agent asset type exists in this integration |
| Yes, from a dataset artifact represented as a Storage Container asset | Yes, to a resultset artifact represented as a Storage Container asset | Not available, no AI Agent asset type exists in this integration | |
| Snowflake Cortex AI | Yes, from an already ingested Snowflake table, found using Snowflake's built-in lineage tracking, and requires Snowflake Enterprise Edition or higher | Not supported, the AI Model Deployment asset exists but does not have an output data relation | Yes, automatic for Cortex Analyst tools, the AI Agent Tool asset links to an already ingested Semantic View asset |
For details on how these relations are created, go to AI model traceability: automatic linking of AI Governance assets.
Operational trust metrics
Operational trust metrics track how an AI agent performs in production over time, using judge pass/fail rates and token consumption metrics.
Like attributes, these metrics are stored internally. Unlike attributes, they aren't accessible through the Collibra Platform APIs. They power the Agents dashboard, the Quality tab on AI Agent assets, and the Operational Health theme on the AI Trust Score.
To explore these metrics, in Settings, under AI Command Center, switch on Operational trust for AI agents.
Supported integrations:
| Integration | Operational trust metrics | Token consumption metrics |
|---|---|---|
| AWS Bedrock AI | Yes, pass/fail/error counts per AI Endpoint (agents) | Yes, per AI Endpoint (agents) and AI Model Deployment (models) |
| Azure AI Foundry | Yes, daily trust/quality score from Azure's continuous evaluation, only for agents built through Azure's Agents API on Foundry project resources | Yes, per AI Model Deployment, for every deployed model in a Foundry project |
| Databricks AI | Yes, pass/fail/no-assessment counts, Knowledge Assistant and Multi-Agent Supervisor agents only | Yes, Multi-Agent Supervisor agents only |
| Gemini Enterprise Agent Platform | Yes, pass/fail counts post automatically once synced | Yes, requires the agent's own code to self-instrument a custom metric |
For more information, go to Operational trust for AI agents and AI monitoring for AI agents.
For the specific metrics, coverage, and limitations for each integration, go to that integration's own documentation.