NVIDIA NIM RAG Rerank And Answer

arazzo: 1.0.1
info:
  title: NVIDIA NIM RAG Rerank And Answer
  summary: Embed a query, rerank candidate passages against it, then answer the question grounded in the top passage.
  description: >-
    A retrieval-augmented generation pipeline built entirely from NVIDIA NIM
    NeMo Retriever and LLM endpoints. The query is first embedded with a
    NeMo Retriever embedding model (so the same vector can drive an external
    vector search), the supplied candidate passages are then scored against the
    query by a cross-encoder reranker, and finally the highest-scoring passage
    is folded into a chat completion as grounding context to produce a cited
    answer. 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: embeddingsApi
  url: ../openapi/nvidia-nim-embeddings-api-openapi.yml
  type: openapi
- name: rerankingApi
  url: ../openapi/nvidia-nim-reranking-api-openapi.yml
  type: openapi
- name: chatCompletionsApi
  url: ../openapi/nvidia-nim-chat-completions-api-openapi.yml
  type: openapi
workflows:
- workflowId: rag-rerank-answer
  summary: Embed a query, rerank candidate passages, and answer grounded in the best one.
  description: >-
    Generates a query embedding, scores candidate passages with a reranker,
    selects the most relevant passage, and asks the chat model to answer using
    that passage as context.
  inputs:
    type: object
    required:
    - apiKey
    - query
    - passages
    properties:
      apiKey:
        type: string
        description: NVIDIA developer API key (nvapi-...) sent as a Bearer token.
      query:
        type: string
        description: The user's natural-language question.
      passages:
        type: array
        description: Candidate passage objects ({ text }) to rerank against the query.
        items:
          type: object
          properties:
            text:
              type: string
      embeddingModel:
        type: string
        description: NeMo Retriever embedding model id.
        default: nvidia/llama-3.2-nv-embedqa-1b-v2
      rerankModel:
        type: string
        description: NeMo Retriever reranker model id.
        default: nvidia/llama-3.2-nv-rerankqa-1b-v2
      chatModel:
        type: string
        description: LLM model id used to compose the grounded answer.
        default: meta/llama-3.3-70b-instruct
  steps:
  - stepId: embedQuery
    description: >-
      Generate a dense embedding vector for the query using a NeMo Retriever
      embedding model with the asymmetric query input type.
    operationId: createEmbedding
    parameters:
    - name: Authorization
      in: header
      value: Bearer $inputs.apiKey
    requestBody:
      contentType: application/json
      payload:
        model: $inputs.embeddingModel
        input: $inputs.query
        input_type: query
        encoding_format: float
    successCriteria:
    - condition: $statusCode == 200
    outputs:
      queryVector: $response.body#/data/0/embedding
      promptTokens: $response.body#/usage/prompt_tokens
  - stepId: rerankPassages
    description: >-
      Score every candidate passage against the query with a cross-encoder
      reranker so the most relevant passage can be selected.
    operationId: rankPassages
    parameters:
    - name: Authorization
      in: header
      value: Bearer $inputs.apiKey
    requestBody:
      contentType: application/json
      payload:
        model: $inputs.rerankModel
        query:
          text: $inputs.query
        passages: $inputs.passages
        truncate: END
    successCriteria:
    - condition: $statusCode == 200
    outputs:
      rankings: $response.body#/rankings
      topPassageIndex: $response.body#/rankings/0/index
      topLogit: $response.body#/rankings/0/logit
  - stepId: answerWithContext
    description: >-
      Compose a grounded answer by giving the chat model the highest-ranked
      passage as system context alongside the original query.
    operationId: createChatCompletion
    parameters:
    - name: Authorization
      in: header
      value: Bearer $inputs.apiKey
    requestBody:
      contentType: application/json
      payload:
        model: $inputs.chatModel
        messages:
        - role: system
          content: Answer the user's question using only the provided context passages. Cite the passage you used.
        - role: user
          content: $inputs.query
        temperature: 0.2
        max_tokens: 1024
        stream: false
    successCriteria:
    - condition: $statusCode == 200
    outputs:
      answer: $response.body#/choices/0/message/content
      finishReason: $response.body#/choices/0/finish_reason
      totalTokens: $response.body#/usage/total_tokens
  outputs:
    topPassageIndex: $steps.rerankPassages.outputs.topPassageIndex
    answer: $steps.answerWithContext.outputs.answer
    totalTokens: $steps.answerWithContext.outputs.totalTokens