generated: '2026-06-20'
method: derived
source: >-
  openapi/*-openapi.yml; build.nvidia.com/llms.txt;
  docs.nvidia.com/nim/large-language-models/latest/api-reference.html
notes: >-
  Cross-cutting request/response semantics for the NIM OpenAI-compatible surface, DERIVED
  from the repo OpenAPI and confirmed against the provider's llms.txt. Because NIM tracks
  the OpenAI API contract, most conventions mirror OpenAI's.
authentication:
  style: bearer-api-key
  header: 'Authorization: Bearer nvapi-...'
  key_prefix: nvapi-
  issuance: https://build.nvidia.com/settings
  ref: authentication/nvidia-nim-authentication.yml
base_url:
  hosted: https://integrate.api.nvidia.com/v1
  self_hosted: http://localhost:8000/v1
content_types:
  request: application/json
  responses: [application/json, text/event-stream]
streaming:
  supported: true
  mechanism: "Server-Sent Events (set stream:true); delta chunks terminated by 'data: [DONE]'"
  endpoints: [/v1/chat/completions, /v1/completions]
idempotency:
  supported: false
  notes: Inference calls are non-idempotent generations; no Idempotency-Key header is documented.
pagination:
  style: none
  notes: >-
    Inference endpoints are stateless and return no collections. /v1/models returns a full
    {object:list, data:[...]} array with no pagination params (build.nvidia.com's catalog
    page uses ?page=N but the API /v1/models does not).
field_expansion:
  supported: false
metadata:
  supported: false
request_tracing:
  request_id_header: none
  notes: No documented x-request-id echo on the hosted endpoint.
versioning:
  ref: lifecycle/nvidia-nim-lifecycle.yml
  path_version: /v1
  container_version: semver image tag
error_envelope:
  shape: '{ "error": { "message", "type", "param", "code" } }'
  format: openai-compatible
  ref: errors/nvidia-nim-problem-types.yml
rate_limiting:
  signaled_by: HTTP 429
  free_tier: ~40 RPM on the hosted build.nvidia.com endpoint; 1,000 signup credits
  ref: rate-limits/nvidia-nim-rate-limits.yml
model_selection:
  notes: The target model is selected by the required `model` string in the request body; switching models is a one-line change.