国产大模型 API 选型全景指南：DeepSeek、GLM、Qwen、Kimi 任务分流与统一网关实战

import os
from openai import OpenAI

API_KEY = os.getenv("GPT345_API_KEY", "sk-your-key")
BASE_URL = "https://api.gpt345.com/v1"

client = OpenAI(api_key=API_KEY, base_url=BASE_URL)

class DomesticModelRouter:
    @staticmethod
    def route_request(prompt: str, is_code_task: bool = False, is_long_doc: bool = False) -> str:
        """根据输入特征与业务标记，动态选择性价比最高的模型 ID"""
        # 1. 超长文档（超过 50,000 字符）优先走 Kimi 1M 窗口
        if is_long_doc or len(prompt) > 40000:
            return "kimi-k3"

        # 2. 编程任务分流
        if is_code_task or any(kw in prompt.lower() for kw in ["def ", "class ", "function", "sql", "bug", "leetcode"]):
            if len(prompt) < 1500:
                # 短指令代码补全与单测，走极速低成本通义千问
                return "qwen-2.5-coder-32b"
            else:
                # 复杂架构重构与长逻辑调试，走 DeepSeek V4
                return "deepseek-v4"

        # 3. 结构化 JSON 与中文合规文档，走 GLM-5.3
        if any(kw in prompt for kw in ["json", "表格", "格式化", "报告", "审核", "条款"]):
            return "glm-5.3"

        # 4. 默认高性价比基线模型
        return "deepseek-v4"

    @classmethod
    def execute(cls, prompt: str, system_prompt: str = "你是一个专业的高效 AI 助手。", **kwargs):
        target_model = cls.route_request(prompt)
        print(f"[*] 动态路由判定：该请求分派给 [{target_model}] 处理")
        
        try:
            resp = client.chat.completions.create(
                model=target_model,
                messages=[
                    {"role": "system", "content": system_prompt},
                    {"role": "user", "content": prompt}
                ],
                **kwargs
            )
            return resp.choices[0].message.content
        except Exception as e:
            # 自动降级至备用高可用模型
            fallback = "glm-5.3" if target_model != "glm-5.3" else "deepseek-v4"
            print(f"[!] {target_model} 调用异常: {e}，自动降级至 {fallback}")
            resp = client.chat.completions.create(
                model=fallback,
                messages=[{"role": "system", "content": system_prompt}, {"role": "user", "content": prompt}],
                **kwargs
            )
            return resp.choices[0].message.content

if __name__ == "__main__":
    # 测试案例 1：短代码补全
    q1 = "写一个 Python 函数，使用二分查找从有序列表寻找目标值的索引。"
    print(DomesticModelRouter.execute(q1, is_code_task=True))
    
    # 测试案例 2：结构化审查
    q2 = "请将以下订单描述提取为严格的 JSON 格式：张三昨天在上海订购了3箱有机苹果，实付120元。"
    print(DomesticModelRouter.execute(q2))