Calling the API

The CLEVI gateway's call paths are divided into the OpenAI-compatible specification and the Anthropic-compatible specification. Both use the same API key, and you can call conversational inference models through either one. Embedding and audio/voice paths exist only in the OpenAI-compatible specification; the Anthropic-compatible specification provides three paths: messages, count_tokens, and models.

You enable streaming with stream in the request body; embeddings and some models and paths do not support it. Audio and voice paths follow rules that differ from general inference, such as returning the audio directly and keeping artifacts in CleviDrive for 3 days.

Model IDs and the calls each model allows are on the Models page. Authentication headers and base URLs are on the Authenticating with an API key page.

Choosing a path

Choose a path based on what you want to do and the SDK you use. The paths in the table are relative to each specification's base URL.

OpenAI-compatible paths

OpenAI 호환 규격의 base URL은 https://platform.clevi.net/api/v2/aiservice/openai/v1 입니다. 요청과 응답 본문은 OpenAI 규격을 따릅니다. 아래 표는 추론과 모델 경로를 호스트 기준 전체 경로로 표기하며, 음성·보이스 경로는 뒤의 음성과 보이스 경로 절에 정리했습니다.

A minimal example that calls chat/completions. Replace sk-... with the key issued on the API keys page, and set model to an ID from the Models page.

curl https://platform.clevi.net/api/v2/aiservice/openai/v1/chat/completions \
  -H "X-API-Key: sk-..." \
  -H "Content-Type: application/json" \
  -d '{
    "model": "cip-5.5-im",
    "messages": [{"role": "user", "content": "안녕하세요"}]
  }'

Other paths

The following paths are for diagnostics, legacy compatibility, and relaying.

요청과 응답 필드 전체는 OpenAPI 3.0 문서 https://platform.clevi.net/api/v2/console/openapi/cloud-inference-v1 에서 확인할 수 있습니다. 경로 목록은 이 문서에서 생성되며, 문서는 콘솔에 로그인한 세션으로만 열립니다. 같은 문서로 클라이언트를 생성할 수 있습니다.

Anthropic-compatible paths

Anthropic 호환 규격의 base URL은 https://platform.clevi.net/api/v2/aiservice/anthropic 이며, Anthropic SDK가 요청 경로에 /v1을 직접 추가합니다. 요청과 응답 본문은 Anthropic Messages 규격을 따릅니다.

A minimal example that calls messages.

curl https://platform.clevi.net/api/v2/aiservice/anthropic/v1/messages \
  -H "X-API-Key: sk-..." \
  -H "anthropic-version: 2023-06-01" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "cip-5.5-im",
    "max_tokens": 1024,
    "messages": [{"role": "user", "content": "안녕하세요"}]
  }'

If you omit the anthropic-version header, the gateway uses 2023-06-01. The gateway passes the anthropic-version and anthropic-beta headers through unchanged.

Streaming (SSE)

When you set stream to true in the request body, the server responds with Server-Sent Events. Support varies by path and model as follows.

The following three headers are sent with the first chunk. After that, each chunk sends a data: line and flushes immediately; an event: line is included only when the upstream provides an event name.

• Content-Type: text/event-stream
• Cache-Control: no-cache
• X-Accel-Buffering: no

On the Anthropic-compatible path, each chunk pairs an event line with a data line.

event: content_block_delta
data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"안"}}

chat/completions chunks arrive as data lines, and the final data: [DONE] is passed through unchanged.

data: {"id":"chatcmpl-...","object":"chat.completion.chunk","choices":[...]}

data: [DONE]

If an error occurs during streaming, the same error JSON as a regular call is sent as a single data: line, and the HTTP status code changes accordingly.

Audio and voice paths

The TTS, STT, and Cosa voice management paths are part of the OpenAI-compatible specification, and a few rules differ from general inference. See the Cosa table on the Models page for which models allow which paths.

The following rules apply only to audio and voice paths.