Interacting with AI Studio models

When creating a new workflow or editing an existing one, in the management console, under Create workflow / Edit workflow, select Constructor.

From the AI Studio section on the left side of the screen, drag the AI Studio models element to relevant workflow location in the constructor window.

You can drag and drop steps into areas marked by the dotted line in the workflow schema, such as Add a step, etc.

In the workflow schema window, click the new AI Studio models section to select it.

At the right of the window, on the Settings tab.

1. In the Step name field, enter a name for the workflow step.

2. Select an AI model:

   Text generation

   1. Select:

      • Base model to use YandexGPT Pro or YandexGPT Lite and, in the Folder ID field, select the folder the model will be invoked in.
      • Fine-tuned model to use a fine-tuned model, and specify the model's ID (URI) in the Model URI field.

   2. In the Temperature field, set the model's response variability: the higher the temperature, the less predictable will be the result. The possible values range from 0 to 1.

   3. In the Number of tokens field, limit the maximum allowed number of tokens in the model's response.

   4. Under Context, select:

      • Messages to set the context of your request to the model as a sequence of individual messages in <Message_sender_role>:<Message_text> format using the button.
      • JSON string to set the context of your request to the model in JSON string format.

   5. Optionally, under Reasoning mode, select the reasoning mode status:

      • Unspecified: Not specified.
      • Disabled: Reasoning mode is disabled.
      • Enabled: Reasoning mode is enabled.

   6. Optionally, under Structured output, select response format:

      • None: The model returns a response formatted with Markdown.
      • JSON object: The model returns a response as a JSON object.
      • JSON schema: The model returns a response as a JSON schema set in the JSON schema field.
   Fine-tuned classification

   1. In the Model URI field, enter a fine-tuned model's ID. For more information, see Classifier models based on YandexGPT.
   2. In the Text field, enter the message text.
   Classifcation by prompt

   1. Select the model and the folder it will be invoked in.

   2. In the Task field, enter the text description of the task for the classifier.

   3. In the Text field, enter the message text.

   4. Under Labels, select:

      • List to enter classes, click .
      • JSON and enter classes in JSON format under JSON object.

   5. Optionally, under Sample queries, select:

      • List to enter request examples using .
      • JSON and enter request examples in JSON format under JSON object.

3. Optionally, in the Timeout, ms field, set the maximum execution time for the current step.

4. Optionally, to set a custom retry policy for a step, expand the Retry policy section and click Retry policy. In the form that appears:

   1. Optionally, in the Initial delay, ms field, set the initial value for a delay between step retries.

   2. Optionally, in the Backoff rate field, set the multiplication factor for delay before each step retry.

   3. Optionally, in the Maximum delay, ms field, set the value for a maximum delay between step retries.

   4. In the Errors field, select the errors for which the step will or will not be retried.

      For detailed information about possible errors, see this section.

   5. Optionally, in the Maximum number of retries field, set the maximum number of step retry attempts.

   6. In the Error selection mode field, select:

      • INCLUDE: Retry executing a step when errors specified in the Errors field occur.
      • EXCLUDE: To retry executing the step on any errors other than those specified in the Errors field.

   If you want to delete a retry policy you created for the step earlier, click and select Delete in the Retry policy row.

   If no custom retry policy is configured for a step, the retry policy set for the whole workflow will apply.

Optionally, navigate to the Input tab and set a jq template to filter the workflow state fed into the step.

Optionally, navigate to the Output tab and set a jq template to filter the step outputs added into the workflow state.

Optionally, add an error transition rule for the step you are creating to handle errors you may get during this step.