Tensor Cortex Gateway API reference

Base URL and authentication

Every request uses the installation-specific HTTPS base URL and a scoped, revocable Tensor Cortex service-account key. The key resolves one application and its allowed logical profiles before the Gateway reads product data or calls a provider.

TC_API_BASE_URL=https://ai.company.example
TC_API_KEY=<deployment-issued-scoped-key>

Example-domain rule: ai.company.example is reserved for documentation. It is not a Tensor Cortex service and will not execute the request.

Endpoints

Chat-completions request

The model field is a logical Tensor Cortex profile such as automatic, private, or coding. It is not a physical provider model ID. Physical provider, deployment, and credential selection remain behind the policy and routing boundary.

curl "$TC_API_BASE_URL/v1/chat/completions" \
  -H "Authorization: Bearer $TC_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "automatic",
    "messages": [
      {"role": "user", "content": "Summarize this incident."}
    ],
    "stream": false
  }'

Request fields

Unknown fields and unsupported values fail before provider execution. A caller may provide trusted metadata headers only when its key has the corresponding scope; caller-supplied application, data-class, request-group, or end-user values never grant authorization or weaken policy.

Response contract

A successful non-streaming response uses the OpenAI-compatible chat-completion envelope. The returned model remains the logical profile; the physical provider and deployment are not exposed on this surface.

{
  "id": "chatcmpl_tc_example",
  "object": "chat.completion",
  "created": 1784246400,
  "model": "automatic",
  "choices": [
    {
      "index": 0,
      "message": {"role": "assistant", "content": "..."},
      "finish_reason": "stop"
    }
  ]
}

Response metadata

Safe error behavior

Error bodies use an OpenAI-compatible error envelope and carry the same request ID. Safe messages exclude prompt content, provider response bodies, credentials, candidate names, and physical model identity.

Compatibility limits

• The committed contract covers GET /v1/models and POST /v1/chat/completions; it is not a claim of full OpenAI API endpoint or extension parity.
• Text chat, streaming, function tools, tool calls, normalized finish reasons, and optional streaming usage are in scope. Provider-specific extensions and uncommon multimodal parts require explicit compatibility evidence.
• Fallback is permitted only inside the same policy-approved candidate set and never after the first client-visible streaming event.
• When effective policy requires durable audit before delivery or encrypted content capture, streaming fails before provider execution and the client must retry with stream=false. The safe error code is streaming_requires_durable_audit.
• Gateway responses are not cached. A repeated request performs a new policy, route, and provider execution.

Privacy and execution boundary

Gateway prompts and responses are not durably stored by default and are not response-cached. Metadata-only audit records identity and resource references, profile and policy versions, data class, route and fallback decisions, token counts, latency, health, and estimated cost—not prompt or response bodies.

An approved deployment controls provider credentials and egress. Prompt content is sent only after authentication, classification, policy enforcement, and route approval. Provider terms, retention, training, region, and subprocessor status remain deployment-specific facts; a provider brand alone does not establish them.