Replicate Search for a Model and Run a Prediction

arazzo: 1.0.1
info:
  title: Replicate Search for a Model and Run a Prediction
  summary: Search public models by query, run a prediction on the top match's latest version, and poll it.
  description: >-
    Discovery-driven inference. The workflow searches public models with a free
    text query, takes the top result's latest version, runs a prediction
    against that version, and polls the prediction until it reaches a terminal
    state. This is useful for agents that resolve a model from a description
    rather than a known ID. 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: search-model-and-predict
  summary: Search public models and run a prediction on the top match's latest version.
  description: >-
    Runs a model search, selects the latest version of the first result, submits
    a prediction against it, and polls the prediction until it succeeds, fails,
    or is canceled.
  inputs:
    type: object
    required:
    - apiToken
    - query
    - input
    properties:
      apiToken:
        type: string
        description: Replicate API token used as a Bearer credential.
      query:
        type: string
        description: The free text search query for public models.
      input:
        type: object
        description: The model's input as a JSON object for the prediction.
  steps:
  - stepId: searchModels
    description: >-
      Search public models with the supplied query and capture the latest
      version ID of the top result. The request body is a plain text query.
    operationId: models.search
    parameters:
    - name: Authorization
      in: header
      value: Bearer $inputs.apiToken
    requestBody:
      contentType: text/plain
      payload: $inputs.query
    successCriteria:
    - condition: $statusCode == 200
    outputs:
      topModelName: $response.body#/results/0/name
      topVersionId: $response.body#/results/0/latest_version/id
  - stepId: createPrediction
    description: >-
      Create a prediction for the top result's latest version with the supplied
      input.
    operationId: predictions.create
    parameters:
    - name: Authorization
      in: header
      value: Bearer $inputs.apiToken
    requestBody:
      contentType: application/json
      payload:
        version: $steps.searchModels.outputs.topVersionId
        input: $inputs.input
    successCriteria:
    - condition: $statusCode == 201
    outputs:
      predictionId: $response.body#/id
  - stepId: getPrediction
    description: >-
      Retrieve the prediction state, repeating via 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: predictionDone
      type: end
      criteria:
      - context: $response.body
        condition: $.status == "succeeded" || $.status == "failed" || $.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:
    topModelName: $steps.searchModels.outputs.topModelName
    topVersionId: $steps.searchModels.outputs.topVersionId
    predictionId: $steps.createPrediction.outputs.predictionId
    status: $steps.getPrediction.outputs.status
    output: $steps.getPrediction.outputs.output