Create Chat Completion (Streaming)

Official Documentation Reference: OpenAI Chat Create

Endpoint Description

Given a prompt, the model will return one or more predicted chat completions. This endpoint supports both standard and streaming (SSE) responses, and can also return the log probabilities of alternative tokens at each position.

Request Details

• Method: POST
• Endpoint: https://api.codingplanx.ai/v1/chat/completions

Headers

Request Body

Request Example

{
  "model": "gpt-5-mini",
  "max_tokens": 1000,
  "messages": [
    {
      "role": "system",
      "content": "You are a helpful assistant."
    },
    {
      "role": "user",
      "content": "你好"
    }
  ],
  "temperature": 1.0,
  "stream": true,
  "stream_options": {
    "include_usage": true
  }
}

---

Response Details

• Content-Type: application/json (Non-streaming) or text/event-stream (When stream: true)

Response Parameters (Non-streaming Structure)

Response Example (Non-streaming / Aggregated Streaming Result)

{
    "id": "chatcmpl-123",
    "object": "chat.completion",
    "created": 1677652288,
    "choices": [
        {
            "index": 0,
            "message": {
                "role": "assistant",
                "content": "\r
\r
Hello there, how may I assist you today?"
            },
            "finish_reason": "stop"
        }
    ],
    "usage": {
        "prompt_tokens": 9,
        "completion_tokens": 12,
        "total_tokens": 21
    }
}

(Note: If stream: true is specified in the request, the response body will be a stream of Server-Sent Events (SSE) separated by newlines, with each line starting with data: {...} and finally terminating with data: [DONE].)

---

FAQs (Frequently Asked Questions)

Q1: How do I parse a streaming response when stream: true is enabled? A: When stream: true is enabled, the API returns data via the Server-Sent Events (SSE) protocol. The client must read the response stream line by line, locate the strings starting with the data: prefix, and parse the subsequent text into JSON. Please note that the final message of the data stream is always data: [DONE], which signals the end of the stream. Ensure your parser filters out this termination marker.

Q2: Why does my request hang or return an infinite blank stream after setting response_format to JSON? A: When you enable JSON mode by using { "type": "json_object" }, you must explicitly instruct the model to "output in JSON format" within the messages array (typically in the system prompt). Without this explicit text instruction, the model may get stuck in a loop generating endless whitespace characters until it triggers the max_tokens limit.

Q3: Can I adjust both temperature and top_p parameters at the same time? A: The official documentation strongly advises against adjusting both parameters simultaneously. Both are mechanisms designed to control the randomness of the model's output. If you want a more stable and predictable output, lower the temperature or decrease top_p; if you desire a more creative response, increase one of them. Modifying both at the same time can lead to unpredictable and suboptimal results.

Q4: What is the purpose of the extra_body.enable_thinking parameter? A: This is a specialized extension parameter for this endpoint. When passing {"enable_thinking": true}, if the backend model being used supports "Thinking Mode" or "Chain of Thought (CoT)" (such as certain reasoning-enhanced LLMs), the model will initiate a deep thinking process before outputting the final response. If the current model does not support this feature, the parameter is typically ignored.

Q5: Why is the finish_reason in the response length instead of stop? A: A finish_reason of length indicates that the model was forcefully truncated before it could generate a complete response. This usually occurs because your max_tokens parameter is set too low, or the combined total of input tokens and generated tokens has exceeded the maximum context window limit allowed by the current model. To resolve this, try increasing the max_tokens value.