生成结构化输出

使用 OpenAI Responses API

James Chapman

AI Curriculum Manager, DataCamp

接下来……

第3章_概览.jpg

结构化的必要性

语言应用1.jpg

结构化的必要性

语言应用2.jpg

结构化的必要性

语言应用3.jpg

第1部分：定义输出模式

from pydantic import BaseModel


class QuizResult(BaseModel):
    score: int
    passed: bool
    feedback: str

第1部分：定义输出模式

from pydantic import BaseModel, Field

class QuizResult(BaseModel):
    score: int = Field(description="10题中答对的数量")
    passed: bool = Field(description="若得分≥7则为 True")
    feedback: str = Field(
        description="带具体改进建议的鼓励性反馈"
    )

第2部分：发送请求

response = client.responses.parse(

    model="gpt-5.4-mini",

    instructions="You are a Spanish vocabulary tutor. Grade the student's quiz
    answers. Grade the quiz with 2 points per correct answer.",
    input="""1. casa = house
             2. perro = dog
             3. gato = car
             4. libro = book
             5. agua = water""",

    text_format=QuizResult

)

第3部分：提取结果

result = response.output_parsed

print(f"Score: {result.score}/10")
print(f"Passed: {result.passed}")
print(f"Feedback: {result.feedback}")

Score: 8/10
Passed: True
Feedback: Great job - you scored 8/10 (4/5 correct). The only mistake was #3: 'gato'
means 'cat', not 'car' (Spanish for 'car' is 'coche' or 'carro'). Tip: review common
animal vocabulary with flashcards and short quizzes to reinforce recall.

更复杂的数据结构

class Mistake(BaseModel):
    word: str = Field(description="题中错误的西语单词")
    student_answer: str = Field(description="学生填写的答案")
    correct_answer: str = Field(description="正确译法")


class DetailedQuizResult(BaseModel):
    score: int = Field(description="10题中答对的数量")
    passed: bool = Field(description="若得分≥7则为 True")
    feedback: str = Field(description="带具体建议的鼓励性反馈")

    mistakes: list[Mistake] = Field(description="错题列表")

response = client.responses.parse(
    model="gpt-5.4-mini",
    instructions="You are a Spanish vocabulary tutor. Grade the student's quiz
    answers. Grade the quiz with 2 points per correct answer.",
    input="""1. casa = house
             2. perro = dog
             3. gato = car
             4. libro = library
             5. agua = water""",

    text_format=DetailedQuizResult

)

result = response.output_parsed
print(f"Score: {result.score}/10")
print(f"Passed: {result.passed}")

for mistake in result.mistakes:
    print(f"{mistake.word}: '{mistake.student_answer}' -> '{mistake.correct_answer}'")

Score: 6/10
Passed: False
gato: 'car' -> 'cat'
libro: 'library' -> 'book'

小结

from pydantic import BaseModel, Field

class QuizResult(BaseModel):
    score: int = Field(...)
    passed: bool = Field(...)
    feedback: str = Field(...)

result = response.output_parsed

print(f"Score: {result.score}/10")
print(f"Passed: {result.passed}")
print(f"Feedback: {result.feedback}")

response = client.responses.parse(
    model="gpt-5.4-mini",
    instructions="...",
    input="...",

    text_format=QuizResult

)

Vamos praticar!

使用 OpenAI Responses API