generated: '2026-08-26'
method: derived
source: openapi/multiverse-computing-compactifai-openapi.yml
also_searched:
- https://docs.compactif.ai/openai-compatibility/
- https://docs.compactif.ai/api_reference/
- https://docs.compactif.ai/error-handling/
- https://docs.compactif.ai/authentication/
- https://docs.compactif.ai/faq/
api_style: REST/JSON, OpenAI-compatible
base_urls:
- url: https://api.compactif.ai/v1
  scope: global
  note: Routes each request to whichever region is fastest at that moment.
- url: https://api-eu.compactif.ai/v1
  scope: EU data residency
- url: https://api-us.compactif.ai/v1
  scope: US data residency
  source: https://docs.compactif.ai/faq/
authentication:
  style: bearer-api-key
  header: Authorization
  format: Bearer <API key>
  scopes: none
  cross_ref: authentication/multiverse-computing-authentication.yml
idempotency:
  supported: false
  header: null
  scope: null
  retention: null
  evidence:
  - No Idempotency-Key (or equivalent) header appears anywhere in the 13-operation OpenAPI.
  - The string "idempoten" does not occur in the OpenAPI document or in any page of
    docs.compactif.ai.
  mitigations:
  - name: seed
    available: false
    note: 'Unlike OpenAI''s chat completions, ChatCompletionRequest here declares no `seed`
      parameter, so even deterministic-sampling replay is unavailable.'
  - name: client-supplied correlation id
    available: true
    header: x-request-id
    note: 'x-request-id is an optional REQUEST header on 11 of the 13 operations, so a
      client can supply its own correlation value — but nothing in the contract says the
      server deduplicates on it. It is a tracing identifier, not an idempotency key.'
  consequence: 'A retried POST /v1/chat/completions, /v1/completions, /v1/responses or
    /v1/batches is billed and executed again. For an agent, the safe posture is
    at-most-once: do not blind-retry a write, reconcile with GET /v1/batches or
    GET /v1/responses/{response_id} first.'
pagination:
  supported: true
  style: cursor
  operations:
  - get_batches_v1_batches_get
  request_params:
  - name: after
    in: query
    required: false
    description: 'A cursor for pagination: the object id after which to start.'
  - name: limit
    in: query
    required: false
    minimum: 1
    maximum: 100
    description: Maximum number of items to return.
  response_envelope:
    schema: BatchList
    fields:
    - object
    - data
    - has_more
    - first_id
    - last_id
  note: 'Only the batches list is paginated. GET /v1/models returns the full catalog with
    no cursor and no limit.'
field_expansion:
  supported: false
  note: No expand/fields/include parameter on any operation.
metadata:
  supported: true
  where: BatchCreateParams.metadata and Batch.metadata
  note: Free-form caller metadata is supported on batch jobs only.
request_tracing:
  request_header: x-request-id
  response_header: x-request-id
  additional_response_headers:
  - x-served-by-region
  - x-envoy-upstream-service-time
  observed:
    url: https://api.compactif.ai/v1/models
    status: 401
    x-served-by-region: us-east-1
  note: 'x-request-id is both accepted on the request (optional, declared on 11 operations)
    and returned on every response. x-app-id is a second optional request header on the
    four inference operations. Neither is documented on docs.compactif.ai.'
versioning:
  style: path-prefix
  current: v1
  policy_published: false
  cross_ref: lifecycle/multiverse-computing-lifecycle.yml
error_envelope:
  primary: '{"error": {"type", "message", "param", "code"}}'
  secondary: '{"detail": ...}'
  rfc9457: false
  cross_ref: errors/multiverse-computing-problem-types.yml
rate_limit_signaling:
  documented: false
  headers: []
  status_on_exhaustion: undocumented
  cross_ref: rate-limits/multiverse-computing-rate-limits.yml
streaming:
  supported: true
  transport: Server-Sent Events
  parameter: 'stream: true'
  operations:
  - chat_completions_v1_chat_completions_post
  - complete_prompt_v1_completions_post
  - post_responses_v1_responses_post
  - create_transcription_v1_audio_transcriptions_post
  source: https://docs.compactif.ai/faq/
content_types:
  json: application/json
  uploads: multipart/form-data (POST /v1/files, POST /v1/audio/transcriptions)
  binary_download: GET /v1/files/{file_id}/content returns raw bytes
dry_run_mode:
  supported: false
  note: No preview/validate/dry-run parameter or endpoint exists. There is also no
    tokenizer or cost-estimate endpoint, so an agent cannot price a call before making it.
reversibility:
  grade: documented
  applicable: true
  summary: 'One real reversal path exists — batch cancellation — and no window is stated
    for it. Every other write on this API is an inference call that is billed at the moment
    it is made and cannot be taken back.'
  surfaces:
  - operation: create_batch_v1_batches_post
    method: POST
    path: /v1/batches
    reversal:
      exists: true
      operation: cancel_batch_v1_batches__batch_id__cancel_post
      method: POST
      path: /v1/batches/{batch_id}/cancel
      semantics: 'Requests cancellation of an in-progress batch; returns the updated Batch,
        usually with status "cancelling" and cancelling_at set. The Batch object also
        carries cancelled_at, so the terminal cancelled state is observable.'
    window:
      stated: false
      value: null
      evidence: 'The spec says cancellation applies to an "in-progress" batch, which bounds
        it by job state rather than by time. Neither docs.compactif.ai nor the OpenAPI
        states how long cancellation remains possible, how long the batch sits in
        "cancelling", or whether already-completed requests inside the batch are still
        billed. BatchCreateParams.completion_window and Batch.expires_at exist but describe
        when the job must FINISH, not how long it can be cancelled.'
      source: openapi/multiverse-computing-compactifai-openapi.yml
    grade: documented
  - operation: post_files_v1_files_post
    method: POST
    path: /v1/files
    reversal:
      exists: false
      note: 'There is no DELETE /v1/files/{file_id}. The API can upload a file and read it
        back, but publishes no way to remove one. OutputExpiresAfter /
        BatchCreateParams.output_expires_after governs expiry of batch OUTPUT files only.'
    window:
      stated: false
    grade: none
  - operation: chat_completions_v1_chat_completions_post
    method: POST
    path: /v1/chat/completions
    reversal:
      exists: false
      note: Inference is executed and metered on receipt. No void, no refund, no credit
        path is published.
    grade: none
  - operation: complete_prompt_v1_completions_post
    method: POST
    path: /v1/completions
    reversal:
      exists: false
    grade: none
  - operation: post_responses_v1_responses_post
    method: POST
    path: /v1/responses
    reversal:
      exists: false
      note: 'Responses created with store:true are retrievable via
        GET /v1/responses/{response_id} but there is no delete operation, so a stored
        response cannot be withdrawn through the API.'
    grade: none
  - operation: create_transcription_v1_audio_transcriptions_post
    method: POST
    path: /v1/audio/transcriptions
    reversal:
      exists: false
    grade: none
  mitigating_context:
  - 'Zero Data Retention is the provider''s stated default: prompts and completions are
    never stored, so "undo" for an inference call is largely moot on the privacy axis —
    but not on the billing axis, since token usage is metered and invoiced monthly.
    Source: https://docs.compactif.ai/data-privacy/'
  agent_guidance: 'Before firing a batch, confirm the caller accepts that cancellation is
    best-effort with no stated window and no stated billing consequence. Treat every
    completion call as irreversible and billed.'
cross_references:
  authentication: authentication/multiverse-computing-authentication.yml
  errors: errors/multiverse-computing-problem-types.yml
  lifecycle: lifecycle/multiverse-computing-lifecycle.yml
  rate_limits: rate-limits/multiverse-computing-rate-limits.yml
  plans: plans/multiverse-computing-plans-pricing.yml