生成追问问题
模型有时可以正确回答子问题，但对整体查询回答错误。这被称为组合性差距 (compositionality gap)1。
我们如何鼓励模型利用子问题的答案来正确生成整体解决方案？
自我提问 (Self-Ask) 是一种使用单个提示词来
- 决定是否需要追问问题
- 生成追问问题
- 回答追问问题
- 回答主要查询
实现¶
import instructor
from openai import OpenAI
from pydantic import BaseModel, Field
client = instructor.from_openai(OpenAI())
class FollowUp(BaseModel):
question: str = Field(description="The follow-up question")
answer: str = Field(description="The answer to the follow-up question")
class Response(BaseModel):
follow_ups_required: bool
follow_ups: list[FollowUp]
final_answer: str
def self_ask(query):
return client.chat.completions.create(
model="gpt-4o",
response_model=Response,
messages=[
{
"role": "system",
"content": f"""Query: {query}
Are follow-up questions needed?
If so, generate follow-up questions, their answers, and then the final answer to the query.
""", # !(1)
},
],
)
if __name__ == "__main__":
query = "Who was president of the U.S. when superconductivity was discovered?"
response = self_ask(query)
print(response.follow_ups_required)
#> True
for follow_up in response.follow_ups:
print(follow_up)
"""
question='When was superconductivity discovered?' answer='Superconductivity was discovered in April 1911.'
"""
"""
question='Who was president of the U.S. in April 1911?' answer='William Howard Taft was the President of the United States in April 1911.'
"""
print(response.final_answer)
"""
William Howard Taft was president of the U.S. when superconductivity was discovered.
"""
- 如果没有
instructor，这个提示通常会作为单样本或少样本提示1来实现，以鼓励通过追问问题进行思考。使用instructor，我们使用零样本提示！