generated: '2026-08-04'
method: searched
source: https://docs.h2o.ai/enterprise-h2ogpte/rest-api, https://docs.h2o.ai/enterprise-h2ogpte/guide/apis, openapi/h2o-ai-h2ogpte-openapi-original.yml
scope: Enterprise h2oGPTe REST API (the H2O MLOps Scoring API is a per-deployment model endpoint and carries none of these conventions)
authentication:
  style: bearer-api-key
  header: 'Authorization: Bearer sk-...'
  scheme_name: bearerAuth
  key_prefix: sk-
  key_types:
  - name: global
    description: >-
      Created without selecting a Collection. Grants full user impersonation and system-wide
      access to every Collection, Document, Chat and setting belonging to that user.
  - name: collection-specific
    description: >-
      Created against one Collection. Permits only chat with that Collection and related
      calls; cannot create, delete, or reach other Collections or Chats.
  docs: https://docs.h2o.ai/enterprise-h2ogpte/guide/apis
  artifact: authentication/h2o-ai-authentication.yml
idempotency:
  supported: false
  evidence: >-
    No Idempotency-Key header, parameter or extension appears anywhere in the 422-operation
    OpenAPI document, and the docs publish no idempotency contract. The only idempotency
    statement in the spec is a per-operation note that DELETE on an admin session returns 204
    whether or not the session existed. Retrying a POST — create_collection,
    create_chat_session, ingest_* — will create a duplicate.
  agent_impact: >-
    An agent driving this API must de-duplicate on its own side (check list_collections /
    list_chat_sessions before creating) because the server offers no replay key.
pagination:
  style: offset-limit
  params:
  - {name: offset, in: query, type: integer, operations: 41}
  - {name: limit, in: query, type: integer, operations: 45}
  response_shape: bare JSON array
  count_operations:
    pattern: separate *_count operations return the total
    examples: [get_collection_count, get_document_count, get_chat_session_count]
  cursors: false
  link_header: false
  note: >-
    Collections list endpoints return a plain array with no envelope, so total counts come
    from the sibling count operations rather than from a wrapper object.
sorting_and_filtering:
  sort_column_param: sort_column
  ascending_param: ascending
  filter_param: filter
  note: Present on the larger list operations; derived from the OpenAPI parameters.
field_expansion:
  supported: false
metadata:
  supported: true
  mechanism: >-
    Collections, Documents and Chat Sessions carry free-form user metadata via dedicated
    operations and properties rather than a universal `metadata` object on every resource.
request_tracing:
  request_id_header: null
  evidence: No X-Request-Id / X-Correlation-Id header appears in the specification.
versioning:
  scheme: uri-path
  current: v1
  base_url: https://h2ogpte.genai.h2o.ai/api/v1
  spec_version: v1.0.0
  product_versioning: semver (h2oGPTe 1.7.x at capture time)
  artifact: lifecycle/h2o-ai-lifecycle.yml
error_envelope:
  format: vendor-json
  schema: EndpointError
  shape: '{"code": <int32>, "message": <string>}'
  rfc9457: false
  artifact: errors/h2o-ai-problem-types.yml
rate_limiting:
  documented_headers: none
  evidence: >-
    No rate-limit header, parameter or response code is declared in the OpenAPI document.
    The h2oGPTe 1.7 changelog announces "per-user API rate limiting and automated key
    deactivation" as an enterprise governance feature, but H2O.ai publishes no numeric
    limits, no retry-after contract and no header names, so no rate-limits artifact is
    asserted here.
  source: https://docs.h2o.ai/enterprise-h2ogpte/changelog
async_model:
  pattern: job-polling
  description: >-
    Long-running work (document ingestion, collection deletion, topic modelling, summarization)
    is started by a create_*_job operation that returns a job handle; the client then polls the
    Jobs operations. There is no webhook or callback surface — see asyncapi/ absence in this repo.
  job_operations: 28
  example_flow: [create_ingest_upload_job, get_job, list_jobs]
content_types:
  request: [application/json, multipart/form-data]
  response: [application/json, text/event-stream]
  streaming: >-
    Chat completions stream over server-sent events; interrupted LLM communication surfaces
    as the ChatError {"error": "..."} envelope.
cross_links:
  authentication: authentication/h2o-ai-authentication.yml
  errors: errors/h2o-ai-problem-types.yml
  lifecycle: lifecycle/h2o-ai-lifecycle.yml
  data_model: data-model/h2o-ai-data-model.yml
  mcp: mcp/h2o-ai-mcp.yml