generated: '2026-08-16'
method: derived
source: openapi/furiosa-predict-v2.yaml, openapi/furiosa-model-repository-v2.yaml
summary: >-
  The Furiosa Model Server data model is a tensor-transport model, not a business-object model.
  There are no persistent resources with identifiers to relate - no accounts, no records, no
  soft-delete. The entity graph is the request/response envelope around a tensor, plus the
  model-repository index that names what can be inferred against. Nothing here has an id prefix
  scheme, because nothing here is stored.
entities:
  - name: inference_request
    description: An inference call - a set of named input tensors plus optionally the outputs to return.
    fields:
      id: string
      parameters: parameters
      inputs: request_input[] (required)
      outputs: request_output[]
    relationships:
      - has_many: request_input
        via: inputs
      - has_many: request_output
        via: outputs
      - has_one: parameters
        via: parameters
  - name: request_input
    description: A single named input tensor - name, shape, datatype and data.
    fields:
      name: string (required)
      shape: integer[] (required)
      datatype: string (required)
      parameters: parameters
      data: tensor_data (required)
    relationships:
      - has_one: tensor_data
        via: data
      - has_one: parameters
        via: parameters
  - name: request_output
    description: A requested output tensor, by name.
    fields:
      name: string (required)
      parameters: parameters
  - name: inference_response
    description: The result of an inference call, echoing model name/version and carrying output tensors.
    fields:
      model_name: string (required)
      model_version: string
      id: string
      parameters: parameters
      outputs: response_output[] (required)
    relationships:
      - has_many: response_output
        via: outputs
      - belongs_to: metadata_model_response
        via: model_name + model_version
  - name: response_output
    description: A single named output tensor returned by the model.
    fields:
      name: string (required)
      shape: integer[] (required)
      datatype: string (required)
      parameters: parameters
      data: tensor_data (required)
    relationships:
      - has_one: tensor_data
        via: data
  - name: tensor_data
    description: The flattened tensor payload - a recursive array of numbers, booleans or strings.
  - name: parameters
    description: An open key/value extension bag attached to requests, inputs, outputs and responses.
  - name: metadata_server_response
    description: Server identity - name, version and supported protocol extensions.
    fields:
      name: string (required)
      version: string (required)
      extensions: string[] (required)
  - name: metadata_model_response
    description: Model identity and tensor contract - platform, versions, inputs and outputs.
    fields:
      name: string (required)
      versions: string[]
      platform: string (required)
      inputs: metadata_tensor[]
      outputs: metadata_tensor[]
    relationships:
      - has_many: metadata_tensor
        via: inputs
      - has_many: metadata_tensor
        via: outputs
  - name: metadata_tensor
    description: The declared name, datatype and shape of one model input or output.
    fields:
      name: string (required)
      datatype: string (required)
      shape: integer[] (required)
  - name: repository_index_request
    description: A request to list the model repository, optionally filtered to ready models.
    fields:
      ready: boolean
  - name: repository_index_response
    description: The repository listing - an array of {name, version, state, reason}.
    relationships:
      - has_many: metadata_model_response
        via: name
        note: Soft join by model name; the index is not a $ref to the model metadata schema.
key_relationships:
  - from: inference_request
    to: metadata_model_response
    type: validated_against
    via: input tensor name/shape/datatype must match the model metadata
    note: >-
      This is the load-bearing relationship for any caller and it is enforced at runtime rather
      than in the schema - inference_request.inputs[].shape is a free integer array, so a mismatch
      surfaces as a 400 inference_error_response, not as a schema validation failure.
  - from: repository_index_response
    to: inference_request
    type: gates
    via: a model must be loaded and ready before /v2/models/{name}/infer will serve it
id_prefixes: []
entity_count: 13
render: null