arazzo: 1.0.1
info:
  title: Replicate Create Prediction and Poll Until Complete
  summary: Run a model version, then poll the prediction until it reaches a terminal state.
  description: >-
    The core asynchronous Replicate pattern. The workflow creates a prediction
    for a specific model version and input, then polls the get prediction
    endpoint until the prediction reaches a terminal status. It branches on the
    returned status so a succeeded prediction surfaces its output while a failed
    or canceled prediction surfaces its error. Every step spells out its request
    inline so the flow can be read and executed without opening the underlying
    OpenAPI description.
  version: 1.0.0
sourceDescriptions:
- name: replicateApi
  url: ../openapi/replicate-openapi.yml
  type: openapi
workflows:
- workflowId: predict-and-poll
  summary: Create a prediction from a model version and poll it until it finishes.
  description: >-
    Submits a prediction request for the supplied model version and input, then
    repeatedly retrieves the prediction until its status is succeeded, failed,
    or canceled, returning the final output or error.
  inputs:
    type: object
    required:
    - apiToken
    - version
    - input
    properties:
      apiToken:
        type: string
        description: Replicate API token used as a Bearer credential.
      version:
        type: string
        description: The ID of the model version that you want to run.
      input:
        type: object
        description: The model's input as a JSON object matching the version's input schema.
      webhook:
        type: string
        description: Optional HTTPS URL to receive webhook callbacks for the prediction.
  steps:
  - stepId: createPrediction
    description: >-
      Create a prediction for the supplied model version and input. Replicate
      returns the prediction in a starting state which must then be polled.
    operationId: predictions.create
    parameters:
    - name: Authorization
      in: header
      value: Bearer $inputs.apiToken
    requestBody:
      contentType: application/json
      payload:
        version: $inputs.version
        input: $inputs.input
        webhook: $inputs.webhook
    successCriteria:
    - condition: $statusCode == 201
    outputs:
      predictionId: $response.body#/id
      initialStatus: $response.body#/status
  - stepId: getPrediction
    description: >-
      Retrieve the current state of the prediction. This step is repeated by the
      retry branch until the prediction reaches a terminal status.
    operationId: predictions.get
    parameters:
    - name: Authorization
      in: header
      value: Bearer $inputs.apiToken
    - name: prediction_id
      in: path
      value: $steps.createPrediction.outputs.predictionId
    successCriteria:
    - condition: $statusCode == 200
    outputs:
      status: $response.body#/status
      output: $response.body#/output
      error: $response.body#/error
    onSuccess:
    - name: predictionSucceeded
      type: end
      criteria:
      - context: $response.body
        condition: $.status == "succeeded"
        type: jsonpath
    - name: predictionFailed
      type: end
      criteria:
      - context: $response.body
        condition: $.status == "failed"
        type: jsonpath
    - name: predictionCanceled
      type: end
      criteria:
      - context: $response.body
        condition: $.status == "canceled"
        type: jsonpath
    - name: keepPolling
      type: retry
      retryAfter: 2
      retryLimit: 60
      stepId: getPrediction
      criteria:
      - context: $response.body
        condition: $.status == "starting" || $.status == "processing"
        type: jsonpath
  outputs:
    predictionId: $steps.createPrediction.outputs.predictionId
    status: $steps.getPrediction.outputs.status
    output: $steps.getPrediction.outputs.output
    error: $steps.getPrediction.outputs.error