考虑更高层次的上下文
我们如何鼓励大型语言模型 (LLM) 思考回答查询所需的任何高层次上下文？回溯提示法分两步鼓励这样做
- 抽象：向大型语言模型 (LLM) 询问一个通用的、更高层次的概念。这通常是特定于主题的。这被称为回溯问题。
- 推理：根据大型语言模型 (LLM) 对抽象问题的回答，询问原始问题。这被称为抽象基础推理。
回溯提示示例
原始问题：当温度和体积增加时，理想气体的压力会发生什么变化？
回溯问题：与这个问题相关的物理概念有哪些？
推理提示：{回溯响应} {原始问题}
请注意，回溯问题也是使用大型语言模型 (LLM) 查询生成的。
回溯提示法已被证明可以提高 PaLM-2L 和 GPT-4 在推理基准上的得分。*
import openai
import instructor
from pydantic import BaseModel
from typing import Iterable, Literal
client = instructor.from_openai(openai.OpenAI())
class Stepback(BaseModel):
original_question: str
abstract_question: str
class Education(BaseModel):
degree: Literal["Bachelors", "Masters", "PhD"]
school: str
topic: str
year: int
class Response(BaseModel):
school: str
def generate_stepback_question():
return client.chat.completions.create(
model="gpt-4o",
response_model=Stepback,
messages=[
{
"role": "user",
"content": f"""
You are an expert at world knowledge. Your task is to step back
and paraphrase a question to a more generic step-back question,
which is easier to answer.
Here are a few examples:
Original Question: Which position did Knox Cunningham hold from
May 1955 to Apr 1956?
Step-back Question: Which positions has Knox Cunningham held in
his career?
Original Question: Who was the spouse of Anna Karina from 1968
to 1974?
Step-back Question: Who were the spouses of Anna Karina?
Original Question: Which team did Thierry Audel play for from
2007 to 2008?
Step-back Question: Which teams did Thierry Audel play for in
his career?
Now, generate the step-back question for the following question:
Estella Leopold went to which school between Aug 1954 and
Nov 1954?
""",
},
],
)
def ask_stepback_question(stepback):
return client.chat.completions.create(
model="gpt-4o",
response_model=Iterable[Education],
messages=[
{"role": "user", "content": stepback.abstract_question},
],
)
def get_final_response(stepback, stepback_response):
return client.chat.completions.create(
model="gpt-4o",
response_model=Response,
messages=[
{
"role": "user",
"content": f"""
Q: {stepback.abstract_question},
A: {stepback_response}
Q: {stepback.original_question}
A:
""",
},
],
)
if __name__ == "__main__":
# Generate the step-back question
stepback = generate_stepback_question()
print(stepback.original_question)
#> Estella Leopold went to which school between Aug 1954 and Nov 1954?
print(stepback.abstract_question)
#> Which schools did Estella Leopold attend in her life?
# Ask the step-back question
stepback_response = ask_stepback_question(stepback)
for item in stepback_response:
print(item)
"""
degree='Bachelors'
school='University of Wisconsin-Madison'
topic='Botany'
year=1948
"""
"""
degree='Masters'
school='University of California, Berkeley'
topic='Botany and Paleobotany'
year=1950
"""
"""
degree='PhD'
school='Yale University'
topic='Botany and Paleobotany'
year=1955
"""
# Ask the original question, appended with context from the stepback response
print(get_final_response(stepback, stepback_response))
#> school='Yale University'