Skip to main content
Dooap Studio
Browse all guides
All guides

AI Models

AI Models are the language, vision, and OCR models that Agent steps and other model-driven steps run on. Studio ships with global models that are available to every tenant, and Tenant Admins can add their own tenant models when they need a specific deployment, provider, or set of credentials.

Where to Find AI Models

Open AI Models from the Agents group in the left navigation. The page is visible to Tenant Admins. It lists the models available to the tenant — both global models (read-only) and any tenant models you have added — and a form for configuring the selected model.

Global vs Tenant Models

  • Global models are provided centrally and can be selected in steps as-is. You cannot edit or delete them.
  • Tenant models are ones you add for your tenant. Use these when you want to point at your own deployment, use your own API credentials, or expose a model that is not available globally.

Model Types

Each model has a type that determines where it can be used:

  • LLM — large language models for text reasoning and generation in Agent steps.
  • VLM — vision language models that can also interpret images.
  • OCR — optical character recognition for reading text from documents and images.

Adding a Tenant Model

  1. Go to AI Models and select New.
  2. Fill in the model details:
    • Model Key — a unique identifier (lowercase letters, numbers, and dashes). It cannot be changed after the model is created.
    • Name — the display name shown when selecting a model in a step.
    • Model Type — LLM, VLM, or OCR.
    • Notes — an optional description.
    • Reasoning Levels and Default Level — the reasoning levels Agent Steps can choose for this model (see Reasoning Levels). Not shown for OCR models.
  3. Fill in the deployment details:
    • Deployment Type — the provider or service, such as Azure OpenAI, OpenAI, Azure Document Intelligence, Azure Content Understanding, Ollama, or Datalab.
    • Provider Model Name — the model identifier at the provider, for example gpt-4o.
    • Deployment Name — the provider's deployment identifier (used by providers such as Azure OpenAI).
    • Endpoint — the service endpoint URL.
    • Access Key — the API key or credential. Leave empty when the deployment authenticates with a managed identity.
    • API Version — the provider API version where one is required.
    • API — for Azure OpenAI and OpenAI deployments, whether Studio calls the model through Chat Completions or the Responses API (see Choosing the API).
    • Configuration — optional additional settings as JSON.
  4. Select Test connection to verify the deployment works the way an Agent Step will call it (see Testing a Deployment).
  5. Select Save.

You can reference Variables and Secrets in credential fields rather than pasting raw values — for example {{SECRET.OPENAI_API_KEY}}.

Reasoning Levels

Reasoning models let you trade answer quality against speed and token usage. Reasoning Levels lists the levels this model accepts — None, Minimal, Low, Medium, High, Extra high, and Max — and Agent Steps that use the model can then pick one of the checked levels in their Reasoning Level setting. The Default Level is used when a step keeps its setting on Default, and when the model is a fallback model that doesn't offer the step's level.

Check only the levels the provider documents for the model, because a level the model rejects makes the step fail. Supported levels differ between models, and sometimes between APIs of the same model:

Model (examples)Levels
GPT-5.4None, Low, Medium, High, Extra high
gpt-6-lunaNone, Low, Medium, High, Extra high (Max only on the Responses API)
Claude Sonnet 5 and laterLow, Medium, High, Extra high, Max
grok-4.3None, Minimal, Low, Medium, High
gpt-4o and other non-reasoning modelsNo levels

Leave all levels unchecked for models without a reasoning setting. Their steps then don't show the option, and Studio sends no reasoning setting to the provider. Global models are configured by system administrators; their levels are shown read-only on the global model's page.

Testing a Deployment

Test connection sends a short question to the deployment through the same API and settings an Agent Step uses. Each request offers the model a small test tool (get_capital_city), and the model has to call it before it answers, so the test also proves that the deployment accepts tool calls.

When the model has Reasoning Levels checked, Studio sends one request for each checked level, using the form's current settings even before you save. The result lists every level with its answer or error. A level fails when the provider rejects the request, the model answers without calling the tool, or the request takes longer than 90 seconds. Connection succeeded means every level worked. If some levels fail, the summary names them. Uncheck those levels, or change the API, before steps depend on them. When no level is checked, Studio sends a single request without a reasoning setting.

The result also shows which API Studio used and the endpoint it resolved.

Choosing the API

Azure OpenAI and OpenAI deployments can be called through two APIs: Chat Completions and Responses. The API option chooses which one:

  • Default uses Chat Completions, except for Codex models, which use Responses. The option shows which one applies.
  • Chat Completions or Responses always uses that API.

The choice is saved as apiOperation in the deployment's Configuration, for example {"apiOperation": "responses"}.

The endpoint URL does not choose the API. An Azure AI Foundry target URI ending in /openai/v1/responses still uses Chat Completions unless API is set to Responses, and the form shows a hint when this is the case. Some newer models need the Responses API for some features. For example, gpt-6-luna accepts tools together with a reasoning level only on the Responses API. On Chat Completions, an Agent Step that uses tools fails unless its reasoning level is None. Test connection reports those levels as failed, so you can switch API to Responses and test again before saving.

How Steps Use Models

Agent steps and other model-driven steps select a model from the list of models available to the tenant. When you promote an Action between environments, the model it referenced in the source tenant may not exist by the same name in the target tenant; the import wizard lets you map it to a model that does. See Promoting from UAT to Production.

Things to Know

  • Model exports do not include access keys. After importing models into another tenant, re-enter their credentials.
  • Deleting a tenant model affects any step that selects it, so check usage before removing one.
  • Use Test connection after changing endpoints, credentials, the API, or reasoning levels, to confirm that the deployment accepts tool calls at every checked level before a real run depends on it.