|
ingFang SC", Cambria, Cochin, Georgia, Times, "Times New Roman", serif;">今天OpenAI的API新增加了一个功能，支持结构化输出，通俗的来说API的输出支持稳定的JSON模式。ingFang SC", Cambria, Cochin, Georgia, Times, "Times New Roman", serif;">为什么强调稳定呢？其实在去年的DevDay，OpenAI就介绍了JSON models，尽管它提高了模型输出是JSON的可靠性，但它不能百分之百确保输出是JSON。而今天提出的结构化输出终于确保API输出是JSON模式了。ingFang SC", Cambria, Cochin, Georgia, Times, "Times New Roman", serif;">其实这个功能非常非常有用，对于调用API的应用程序来说，从非结构化输入生成结构化输出是最核心的用例，想当年我们在调用OpenAI接口的时候，通过prompt让接口输出JSON，可有的时候行，有的时候又不行，根本找不到规律，现在这个问题终于解决了！ingFang SC", Cambria, Cochin, Georgia, Times, "Times New Roman", serif;">OpenAI经过测试，gpt-4o-2024-08-06支持结构化输出，其输出JSON的稳定性是100%，而gpt-4-0613得分不到40%，如下图： 
ingFang SC", Cambria, Cochin, Georgia, Times, "Times New Roman", serif;">说了这么多，如何使用API的结构化输出呢？有两种模式。ingFang SC", Cambria, Cochin, Georgia, Times, "Times New Roman", serif;">1：Function callingingFang SC", Cambria, Cochin, Georgia, Times, "Times New Roman", serif;">在函数定义中设置 strict: true，即可通过tools进行结构化输出，如下图：
ingFang SC", Cambria, Cochin, Georgia, Times, "Times New Roman", serif;">其输出大概如下：
ingFang SC", Cambria, Cochin, Georgia, Times, "Times New Roman", serif;">2：response_format 参数ingFang SC", Cambria, Cochin, Georgia, Times, "Times New Roman", serif;">上述Function calling的方式个人觉得非常复杂，所以也可以在API中使用新增的response_format 参数：POST/v1/chat/completions{"model":"gpt-4o-2024-08-06","messages":[{"role":"system","content":"Youareahelpfulmathtutor."},{"role":"user","content":"solve8x+31=2"}],"response_format":{"type":"json_schema","json_schema":{"name":"math_response","strict":true,"schema":{"type":"object","properties":{"steps":{"type":"array","items":{"type":"object","properties":{"explanation":{"type":"string"},"output":{"type":"string"}},"required":["explanation","output"],"additionalProperties":false}},"final_answer":{"type":"string"}},"required":["steps","final_answer"],"additionalProperties":false}}}}输出如下： {"steps":[{"explanation":"Subtract31frombothsidestoisolatethetermwithx.","output":"8x+31-31=2-31"},{"explanation":"Thissimplifiesto8x=-29.","output":"8x=-29"},{"explanation":"Dividebothsidesby8tosolveforx.","output":"x=-29/8"}],"final_answer":"x=-29/8"}虽然API request看上去很复杂，比如Python官方包做了封装，实现上述功能的代码如下： from pydantic import BaseModelfrom openai import OpenAI
class Step(BaseModel):explanation: stroutput: str
class MathResponse(BaseModel):steps: list[Step]final_answer: str
client = OpenAI()completion = client.beta.chat.completions.parse(model="gpt-4o-2024-08-06",messages=[{"role": "system", "content": "You are a helpful math tutor."},{"role": "user", "content": "solve 8x + 31 = 2"},],response_format=MathResponse,)
message = completion.choices[0].messageif message.parsed:print(message.parsed.steps)print(message.parsed.final_answer)else:print(message.refusal)
此外OpenAI特别强调了结构化输出功能是安全的，必须符合现有的安全策略，所以特别在API中新增加了一个输出字段refusal，用于描述是不符合JSON还是不安全！ 这篇文章描述了如何使用结构化输出功能，其实OpenAI也介绍了如何实现该功能，这个有机会再讲，对于理解模型训练很有帮助！ |
