流式输出

在请求中设置 "stream": true,响应将以 Server-Sent Events(SSE) 逐段返回, 适合聊天界面的打字机效果与长回复场景。

请求

curl https://api.4allapi.com/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $AIPROXY_KEY" \
  -d '{
    "model": "gpt-5.6",
    "messages": [{"role": "user", "content": "写一首关于星空的短诗"}],
    "stream": true,
    "stream_options": {"include_usage": true}
  }'

响应流

每个事件是一行 data: 前缀的 JSON,增量内容在 choices[0].delta.content 中;流结束时发送 data: [DONE]:

data: {"choices":[{"delta":{"content":"夜"}}], ...}
data: {"choices":[{"delta":{"content":"空"}}], ...}
...
data: {"choices":[],"usage":{"prompt_tokens":18,"completion_tokens":56,"total_tokens":74}}
data: [DONE]

SDK 示例(Python)

stream = client.chat.completions.create(
    model="gpt-5.6",
    messages=[{"role": "user", "content": "写一首关于星空的短诗"}],
    stream=True,
)
for chunk in stream:
    delta = chunk.choices[0].delta.content if chunk.choices else None
    if delta:
        print(delta, end="", flush=True)

注意事项

流式连接的读超时建议设置在 60 秒以上,长回复的推理模型可适当加大;
客户端中断连接后,已生成部分仍会按实际用量计费;
代理/网关转发时需关闭响应缓冲(如 nginx 的 proxy_buffering off),否则会失去流式效果。