Hugging Face TGI Inspect and Generate

arazzo: 1.0.1
info:
  title: Hugging Face TGI Inspect and Generate
  summary: Read a TGI server's info, tokenize a prompt to check its length, then generate text.
  description: >-
    A self-hosted Text Generation Inference (TGI) flow. The workflow reads the
    TGI server's info to learn the loaded model and its maximum input length,
    tokenizes the prompt to measure how many tokens it consumes, and then runs a
    non-streaming text generation request against the loaded model. The info and
    tokenize steps ground the generation request in the server's real limits.
    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: textGenerationInferenceApi
  url: ../openapi/hugging-face-text-generation-inference-api.yml
  type: openapi
workflows:
- workflowId: tgi-inspect-and-generate
  summary: Inspect a TGI server, tokenize a prompt, and generate text from it.
  description: >-
    Reads TGI server info, tokenizes the prompt to measure its token count, and
    generates text from the prompt against the loaded model.
  inputs:
    type: object
    required:
    - hfToken
    - prompt
    properties:
      hfToken:
        type: string
        description: Hugging Face access token used as a Bearer credential.
      prompt:
        type: string
        description: The input prompt to tokenize and generate from.
      maxNewTokens:
        type: integer
        description: Maximum number of new tokens to generate.
        default: 128
      temperature:
        type: number
        description: Sampling temperature.
        default: 0.7
  steps:
  - stepId: getServerInfo
    description: >-
      Read the TGI server info to learn the loaded model id and its maximum
      input length before submitting a prompt.
    operationId: getInfo
    successCriteria:
    - condition: $statusCode == 200
    outputs:
      modelId: $response.body#/model_id
      maxInputLength: $response.body#/max_input_length
  - stepId: tokenizePrompt
    description: >-
      Tokenize the prompt to measure how many tokens it consumes against the
      server's input limit.
    operationId: tokenize
    parameters:
    - name: Authorization
      in: header
      value: Bearer $inputs.hfToken
    requestBody:
      contentType: application/json
      payload:
        inputs: $inputs.prompt
        add_special_tokens: true
    successCriteria:
    - condition: $statusCode == 200
    outputs:
      firstTokenId: $response.body#/0/id
  - stepId: generateText
    description: >-
      Generate text from the prompt using the loaded model with details enabled
      so the finish reason and token count are returned.
    operationId: generate
    parameters:
    - name: Authorization
      in: header
      value: Bearer $inputs.hfToken
    requestBody:
      contentType: application/json
      payload:
        inputs: $inputs.prompt
        parameters:
          max_new_tokens: $inputs.maxNewTokens
          temperature: $inputs.temperature
          details: true
          return_full_text: false
    successCriteria:
    - condition: $statusCode == 200
    outputs:
      generatedText: $response.body#/generated_text
      finishReason: $response.body#/details/finish_reason
      generatedTokens: $response.body#/details/generated_tokens
  outputs:
    modelId: $steps.getServerInfo.outputs.modelId
    generatedText: $steps.generateText.outputs.generatedText
    finishReason: $steps.generateText.outputs.finishReason