如何将上下文学习的Completions API输入转换为Chat Completions API格式？

Chat Completions API的输入核心是messages数组，每个元素包含role（可选值：system、user、assistant）和content字段。我们可以把原Prompt拆分到对应角色中，让模型更清晰地理解任务约束与请求：

拆分逻辑 #

• system角色：存放核心规则、任务目标与约束条件，定义模型的行为模式。
• user角色：存放示例数据与当前待处理的输入请求，作为具体任务内容。

完整Python代码示例 #

from openai import OpenAI

client = OpenAI()

response = client.chat.completions.create(
    model="gpt-3.5-turbo",
    messages=[
        {
            "role": "system",
            "content": """你是一个能根据输入序列预判动作的智能助手。执行动作的人是右利手。你只能选择示例中出现过的动作。每个动作都用方括号包裹，例如['Pick', 'cucumber_attachment']或['Cut cucumber_attachment with knife_black']。
请按任意有效顺序完成以下动作：
['RollOut dough']
['Open milk_small']
['Pour milk_small into mixing_bowl_green']
['Pour flour from cup_small into mixing_bowl_green'] 或 ['Scoop flour from cup_small into mixing_bowl_green']
['Scoop mixed contents from mixing_bowl_green into cup_large']
['End']"""
        },
        {
            "role": "user",
            "content": """以下是示例：

Sequence = [['Pick', 'milk_small'], ['Open milk_small'], ['Place', 'milk_small_lid'], ['Pick', 'mixing_bowl_green'], ['Pour milk_small into mixing_bowl_green'], ['Place', 'mixing_bowl_green'], ['Place', 'milk_small'], ['Pick', 'tablespoon'], ['Pick', 'cup_small'], ['Scoop flour from cup_small into mixing_bowl_green'], ['Place', 'tablespoon'], ['Place', 'cup_small'], ['Pick', 'egg_whisk'], ['Pick', 'cup_large'], ['Place', 'cup_large'], ['Pick', 'mixing_bowl_green'], ['Stir contents of mixing_bowl_green'], ['Place', 'egg_whisk'], ['Pick', 'ladle'], ['Scoop mixed contents from mixing_bowl_green into cup_large'], ['Place', 'ladle'], ['Place', 'mixing_bowl_green'], ['Pick', 'rolling_pin'], ['RollOut dough'], ['Place', 'rolling_pin'], ['End']]
Sequence = [['Pick', 'rolling_pin'], ['RollOut dough'], ['Place', 'rolling_pin'], ['Pick', 'tablespoon'], ['Pick', 'cup_small'], ['Scoop flour from cup_small into mixing_bowl_green'], ['Place', 'tablespoon'], ['Place', 'cup_small'], ['Pick', 'milk_small'], ['Open milk_small'], ['Pour milk_small into mixing_bowl_green'], ['Close milk_small_lid'], ['Pick', 'milk_small'], ['Place', 'milk_small'], ['Pick', 'egg_whisk'], ['Stir contents of mixing_bowl_green'], ['Place', 'egg_whisk'], ['Pick', 'ladle'], ['Pick', 'mixing_bowl_green'], ['Place', 'mixing_bowl_green'], ['Pick', 'cup_large'], ['Place', 'cup_large'], ['Scoop mixed contents from mixing_bowl_green into cup_large'], ['Place', 'ladle'], ['End']]

现在，请预判以下输入对应的动作：
Input = [['Pick', 'milk_small']]"""
        }
    ]
)

print(response.choices[0].message.content)

关键说明 #

• 这种拆分方式符合Chat模型的对话设计逻辑，system消息确立模型的行为边界，user消息传递具体任务数据。
• 如果需要多轮交互，可以继续在messages数组中追加assistant（模型回复）和user（新请求）的消息对。

内容的提问来源于stack exchange，提问作者Kong