API Compatibility

UnieAI Runtime exposes an OpenAI-compatible interface. Most OpenAI SDK and HTTP clients need only a different base URL, API key, and model slug. Compatibility depends on the implemented endpoint and the selected base model's upstream capabilities; it is not a guarantee of complete feature parity with every OpenAI API.

Basic configuration

An API key belongs to the personal or organization workspace where it was created and does not grant additional model access.

curl https://your-runtime-domain/v1/chat/completions \
  -H "Authorization: Bearer $UNIEAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"your-model-slug","messages":[{"role":"user","content":"Hello"}]}'

Public endpoints

For some unimplemented POST /v1/* JSON or multipart endpoints, Runtime transparently forwards to the selected model's upstream. A valid key, a model, and a forwardable upstream connection are required.

Chat Completions

Non-standard fields can be accepted, ignored, or rejected upstream. Pass top_k through extra_body in the Python OpenAI SDK:

response = client.chat.completions.create(
    model="your-model-slug",
    messages=[{"role": "user", "content": "Propose three options"}],
    temperature=0.7,
    max_tokens=800,
    extra_body={"top_k": 40},
)

Qwen thinking

chat_template_kwargs.enable_thinking is specific to the Qwen base-model family. It applies to a Qwen base model and to a custom model built from a Qwen base model; family matching is case-insensitive.

{
  "model": "your-qwen-based-model",
  "messages": [{"role":"user","content":"Analyze this step by step"}],
  "chat_template_kwargs": {"enable_thinking": true}
}

• true explicitly enables thinking.
• false explicitly disables it.
• An explicit request value takes precedence over the platform or custom-model default for that request.
• A non-Qwen upstream can ignore or reject the field.

Streaming responses

stream: true uses text/event-stream. Standard text arrives in Chat Completions chunks and ends with data: [DONE].

A custom model can also emit UnieAI metadata chunks, including rag_metadata, rag_iteration_metadata, api_metadata, mcp_metadata, mysql_metadata, skills_metadata, memory_metadata, websearch_metadata, agent_metadata, and guardrail_metadata.

Text-only clients should ignore chunks without choices[].delta.content. Clients rendering citations, tool timelines, or guardrail decisions can parse the relevant metadata and should ignore unknown fields for forward compatibility.

Responses and Anthropic compatibility

Responses API

POST /v1/responses translates instructions, text and image input, function calls and outputs, temperature, top_p, and function tools. Responses built-in tool types without Chat Completions semantics, such as some web_search, local_shell, or custom tools, can be skipped.

Anthropic Messages

POST /v1/messages translates Anthropic messages, content blocks, tools, and tool choice into the internal Chat Completions flow. Stop reasons, usage, and streaming events are mapped back to Anthropic shapes, but provider-specific fields are not guaranteed to survive unchanged.

Response headers

Not every response has every header, particularly errors, unbilled flows, and output guardrail decisions made after streaming starts.

Errors and troubleshooting

{"error":{"code":"unauthorized","message":"Invalid API key","request_id":"req_abc123","details":{}}}

For Model not found, put the slug shown on the custom model's Overview in model, not its display name. Request logs are under Operate → Request logs.

Related documentation

• API Key Management
• Feedback API
• Knowledge Base and Feedback APIs
• Guardrails
• Billing, Wallets, and Metering