Adding Models
Configure custom and non-supported models in AgentMark
Custom model schemas give developers control over non-supported models in AgentMark:
-
Custom Model Integration: Define schemas for models that aren’t natively supported by AgentMark, including proprietary or third-party models
-
User Interface Control: Precisely specify which parameters platform users can modify through the UI, while keeping sensitive or advanced configurations under developer control
-
Cost Customization: Customize the costs for your custom models to enable proper tracking of logs/metrics/traces in AgentMark
This section is only for custom or non-supported models. Built-in models like GPT-4, Claude, and other natively supported models are pre-configured in AgentMark and don’t need schemas defined here.
Pulling Models
To add models to your project, use the pull-models
command:
This interactive command will:
- Show you available providers (like OpenAI, Anthropic)
- Let you select which models you want to use
- Add the appropriate model schemas to your
agentmark.json
file
Adding Custom Model Schemas
Each custom model schema allows you to:
- Set a user-friendly display label
- Configure available settings and their UI controls
- Define parameter constraints and defaults
- Control the order and presentation of settings
Here’s an example of a custom model schema:
Multiple Custom Models
You can define multiple custom models in your schema, each with its own configuration:
Setting Properties
Each setting in a custom model schema can define:
- minimum/maximum: The allowed range for numeric values
- default: The initial value
- multipleOf: Step size for numeric inputs
- label: Display name in the UI
- order: Control the display order in forms
- type: The type of the setting (e.g. “number”, “string”, “boolean”)
- ui: The UI control to use for the setting (e.g. “slider”…more to come)
Custom Model Cost Configuration
You can customize the costs for your custom models to enable proper tracking in AgentMark:
Custom model schemas define the UI and validation rules for non-supported models. The actual implementation details for each model are configured in your provider settings.