About Bring Your Own Model (in preview)
Bring Your Own Model (BYOM) lets you choose which AI model powers each of Collibra's AI features. You can use the models Collibra provides by default, or connect your own model provider and use your own credentials instead. Collibra groups its AI features into three effort tiers, based on how complex the underlying task is. For each tier, you choose either a Collibra-provided model or a model from your own provider to power it.
You can find the AI Model assignments and connect your own models in Settings → Maestro, then select the Models (In preview) tab.
About models and model providers
By default, each effort tier uses a Collibra-provided model. You can connect your own model provider and select your own models to have more control over where inference runs.
Collibra currently supports connecting your own Gemini Enterprise provider. A connection stores the project, endpoint, and credentials Collibra uses to reach your models. A single connection can host more than one model. After you add a connection and a model, you can assign that model to any effort tier in place of a Collibra-provided model.
The Available models table on the Models tab shows every model you can assign, including its provider, connection, model name, model ID, and status:
| Status | Description |
|---|---|
| Collibra managed | This status shows for certified, Collibra-hosted models. You cannot edit Collibra connection details. |
| Online | This status shows for custom models that have a validated connection to Collibra. |
| Model unreachable | This status shows for already configured custom models if the provider no longer supports the model. |
| Provider unreachable | This status shows for custom models if the provider itself is unreachable, for example if the credentials expired |
Important If your project does not have access to a model that a tier needs or a provider is unreachable, requests for that tier return an error. Update your project's model access, or assign a different model to the tier, to resolve the issue.
Effort tiers
An effort tier groups Collibra's AI features by task complexity, so you can assign the right model to each tier instead of configuring every feature individually.
The following table describes each tier and the type of model it works best with:
| Tier | What it is used for | Recommendation |
|---|---|---|
| Low effort | Simple tasks that do not need an expensive model. For example, Data & Analytics or Glossary Semantic Search. | A lightweight, lower-cost model. |
| Medium effort | Most of Collibra's AI features. For example, SQL Assistant for Data Quality, Asset Description Generation, Semantic Mapping Agent, and Picture Descriptions in Diagrams. | A mid-range model, such as a Flash-class or Sonnet-class model. |
| High effort | The most complex features, which need multi-step reasoning. For example, Maestro Agents, Maestro Studio, and DQ AI Rules Proxy | A high-powered model with strong reasoning capabilities. |
If a tier's assigned model is not performing well, assign a higher-grade model to that tier and test the results again. If a tier is already performing well, you can assign a lower-cost model to reduce cost, but there may be a drop in quality.
Certified models
Collibra certifies specific models for specific tiers. A certified model has passed our tests and performs as expected for the feature it supports. We support outputs from certified models through our standard support process.
Right now, only Google Gemini Enterprise models are certified.
Our certified models are:
- Low effort tasks: Gemini 2.58 Flash-Lite
- Medium effort tasks: Gemini 2.5 Flash
- High effort tasks: Gemini 3.8 Flash
FAQ
- Where do I configure my models?
- Go to Settings → Maestro, then select the Models tab.
- Can I use my own model provider for any tier?
- Yes. Each tier can use either a Collibra-provided model or a model from your own provider.
- What happens if I assign a weaker model to a tier?
- The feature still works, but results may be less accurate or less useful.
- What happens if my Gemini Enterprise project does not have access to a required model?
- Requests for that tier return a “Model not available” error until you update your project's model access or assign a different model.
Add a model to start using your own Gemini Enterprise provider.