官网API (科学上网查看)
1、我对该功能的理解
利用gpt的上下文理解能力，在执行方法run_conversation(xx)时，目标锁定在--提取出functions里每个function下required属性对应的值。
而真正的function函数（get_current_weather(xx)）只是对提取出的信息进行了一个json封装，并没有实际功能。
2、举例
1）如本例中functions里只有一个function，该function的"required": ["location"], 即要提取出文本中的location。
2）而 "properties": {
"location": {
"type": "string",
"description": "The city and state, e.g. San Francisco, CA",
},
"unit": {"type": "string", "enum": ["celsius", "fahrenheit"]},
},
properties中关于"location"的描述"description"，我理解就是喂入模型的prompt的一部分，即用户user表达了想让模型替自己干什么。
3）为什么说只是一部分？
因为prompt除了要表述想要AI干什么，还要告诉AI他要从哪里提取，即告诉待处理的信息 messages=[{"role": "user", "content": "What's the weather like in Boston?"}]。
4）这样就给gpt了完整的prompt信息。
3、至于为什么会发送两次gpt调用？
第1次调用gpt目的：即上面第2条-例子中的说明；
第2次调用gpt目的：把第1次提取到信息和function返回的json, 重新组合成对话式的语言，而不只是返回function规定死的json。这样不只是开发人员能看懂，而是任何一个人类都可以看懂。
本例中：
1) function返回的json 如下:
这里location和unit的信息是从message里取的，但message里并没有保存这2个信息，所以这里都是null。数据的真实信息在 message["function_call"]["arguments"]中。
{
weather_info: {
"location": null,
"temperature": "72",
"unit": null,
"forecast": ["sunny", "windy"],
}
}
2) gpt第2次调用重新组装的对话语言：见choices--message--content
{
"choices": [
{
"finish_reason": "stop",
"index": 0,
"message": {
"content": "The current weather in Boston, MA is 72\u00b0F (22\u00b0C). It is sunny and windy.", # 这里的回答把function返回的json值重新组装成了对话语言。而且每次运行回答都会变，但意思都是一样的，也就是同一句话会换个说法，这样就更加符合人类语言的灵活性了。
"role": "assistant"
}
}
],
"created": 1687312724,
"id": "chatcmpl-7ThPI7pl75aEQMV2NvcfNZK051YtH",
"model": "gpt-3.5-turbo-0613",
"object": "chat.completion",
"usage": {
"completion_tokens": 21,
"prompt_tokens": 70,
"total_tokens": 91
}
}
4、代码
（1） 官网api测试代码详解
#!/usr/bin/env python
# -*- coding: UTF-8 -*-
import openai
import json
openai.api_base = '代理地址'
openai.api_key = 'your own apikey'
# Example dummy function hard coded to return the same weather
# In production, this could be your backend API or an external API
def get_current_weather(location, unit="fahrenheit"):
"""Get the current weather in a given location"""
weather_info = {
"location": location,
"temperature": "72",
"unit": unit,
"forecast": ["sunny", "windy"],
}
return json.dumps(weather_info)
# Step 1, send model the user query and what functions it has access to
def run_conversation():
response = openai.ChatCompletion.create(
model="gpt-3.5-turbo-0613",
messages=[{"role": "user", "content": "What's the weather like in Boston?"}],
functions=[
{
"name": "get_current_weather",
"description": "Get the current weather in a given location",
"parameters": {
"type": "object",
"properties": {
"location": {
"type": "string",
"description": "The city and state, e.g. San Francisco, CA",
},
"unit": {"type": "string", "enum": ["celsius", "fahrenheit"]},
},
"required": ["location"],
},
}
],
function_call="auto",
)
message = response["choices"][0]["message"]
print("*"*60)
print(message)
print("*" * 60)
# ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** **
# {
# "content": null,
# "function_call": {
# "arguments": "{\n \"location\": \"Boston, MA\"\n}",
# "name": "get_current_weather"
# },
# "role": "assistant"
# }
# ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** **
# Step 2, check if the model wants to call a function
if message.get("function_call"):
function_name = message["function_call"]["name"]
# Step 3, call the function
# Note: the JSON response from the model may not be valid JSON
function_response = get_current_weather(
location=message.get("location"), # 这里message里并没有保存location和unit的信息，真实信息在 message["function_call"]["arguments"]中
unit=message.get("unit"),
)
# Step 4, send model the info on the function call and function response
second_response = openai.ChatCompletion.create(
model="gpt-3.5-turbo-0613",
messages=[
{"role": "user", "content": "What is the weather like in boston?"},
message,
{
"role": "function",
"name": function_name,
"content": function_response,
},
],
)
return second_response
print(run_conversation())
（2）自定义函数功能
#!/usr/bin/env python
# -*- coding: UTF-8 -*-
import openai
import json
####################################################1、测试function功能
openai.api_base = '代理地址' # Your Azure OpenAI resource's endpoint value.
openai.api_key = 'your own apikey'
import json
def get_tax_info(location):
tax_info = {
"location": location,
"forecast": "缴税地"
}
return json.dumps(tax_info)
# Step 1, send model the user query and what functions it has access to
def run_conversation(content):
response = openai.ChatCompletion.create(
model="gpt-3.5-turbo-0613",
messages=[{"role": "user", "content": content}],
functions=[
{
"name": "get_tax_info",
"description": "提取交税地点",
"parameters": {
"type": "object",
"properties": {
"location": {
"type": "string",
# TODO 这里的描述是prompt的一部分，会影响模型的提取结果。
# 实测，content = "我在安吉开的有共工作室，同时在乐清也经营的有小店，我在这两个地方都要交税。"下，
# 设置"description": "交税地点，比如浙江省-湖州市-安吉县。如果有多个缴税地，则以数组的形式返回,如[浙江省-温州市-乐清市,浙江省-杭州市]，模型提取结果：{'location': '浙江省-湖州市-安吉县,浙江省-温州市-乐清市'}
# 设置"description": "交税地点，比如浙江省-湖州市-安吉县。如果有多个缴税地，则以数组的形式返回,如[浙江省-湖州市-安吉县,浙江省-杭州市]，模型提取结果：{'location': '浙江省-湖州市-安吉县'}
# 设置"description": "交税地点，比如浙江省-湖州市-安吉县。如果有多个缴税地，则以数组的形式返回,如[浙江省-杭州市]，模型提取结果：{'location': '浙江省-湖州市-安吉县'}
# 设置"description": "交税地点，比如浙江省-湖州市-安吉县。如果有多个缴税地，则以数组的形式返回,如[浙江省-温州市-乐清市,浙江省-湖州市-安吉县]，模型提取结果：{'location': '浙江省-湖州市-安吉县,浙江省-温州市-乐清市'}
"description": "交税地点，比如浙江省-湖州市-安吉县。如果有多个缴税地，则以数组的形式返回,如[浙江省-温州市-乐清市,浙江省-杭州市]",
"description": "交税地点，比如浙江省湖州市安吉县",
}
},
"required": ["location"],
},
}
],
function_call="auto",
)
message = response["choices"][0]["message"]
print("*" * 60)
print(message)
print(type(message)) # <class 'openai.openai_object.OpenAIObject'>
print("*" * 60)
# mess_dict = message.to_dict()
# print(mess_dict)
# print(type(mess_dict)) # <class 'dict'>
# print("*" * 60)
# json_str = json.dumps(mess_dict)
# print(json_str)
# print(type(json_str)) # <class 'str'>
print("*" * 60)
arguments_ = message["function_call"]["arguments"]
print(type(arguments_)) # <class 'str'>
print(arguments_) # 这里已经是正常显示的中文了，所以不需要对字符串再次unicode解码
# print(arguments_.encode('unicode-escape').decode('unicode-escape'))
print("*" * 60)
arg_json = json.loads(arguments_) # <class 'dict'>
print(type(arg_json))
print(arg_json)
return message
# # Step 2, check if the model wants to call a function
# if message.get("function_call"):
# function_name = message["function_call"]["name"]
#
# # Step 3, call the function
# # Note: the JSON response from the model may not be valid JSON
# function_response = get_tax_info(
# location=message.get("location")
# )
#
# # Step 4, send model the info on the function call and function response
# second_response = openai.ChatCompletion.create(
# model="gpt-3.5-turbo-0613",
# messages=[
# {"role": "user", "content": content},
# message,
# {
# "role": "function",
# "name": function_name,
# "content": function_response,
# },
# ],
# )
# return second_response
# content = "我在安吉开的有共工作室，同时在乐清也经营的有小店，我在这两个地方都要交税。提取下我的交税地点，并组装成省-市-区的格式，如提取到杭州，则组装返回浙江省-杭州市。"
content = "我在安吉开的有共工作室，同时在乐清也经营的有小店，我在这两个地方都要交税。"
# content = "我在杭州上班，也在这里缴税"
result = run_conversation(content)
####################################################2、测试unicode解码
# str = '''
# {
# "choices": [
# {
# "finish_reason": "stop",
# "index": 0,
# "message": {
# "content": "\u5f88\u62b1\u6b49\uff0c\u6211\u65e0\u6cd5\u63d0\u4f9b\u5173\u4e8e\u5728\u676d\u5dde\u7f34\u7a0e\u7684\u8be6\u7ec6\u4fe1\u606f\u3002\u8bf7\u60a8\u54a8\u8be2\u5f53\u5730\u7a0e\u52a1\u5c40\u6216\u8005\u76f8\u5173\u7a0e\u52a1\u90e8\u95e8\uff0c\u4ed6\u4eec\u5c06\u80fd\u591f\u4e3a\u60a8\u63d0\u4f9b\u6b63\u786e\u7684\u4fe1\u606f\u548c\u6307\u5bfc\u3002",
# "role": "assistant"
# }
# }
# ],
# "created": 1686797348,
# "id": "chatcmpl-7RXKm2DOIKxDtyCkfHF1Gz2xEW7Yi",
# "model": "gpt-3.5-turbo-0613",
# "object": "chat.completion",
# "usage": {
# "completion_tokens": 63,
# "prompt_tokens": 60,
# "total_tokens": 123
# }
# }
# '''
# print('*'*50)
#
# p = "\u65e0\u6cd5\u8bc6\u522b\u5c5e\u6027\u201cphysical_network, network_type\u201d"
# p2 = "\u8d26\u6237\u8bbe\u5907"
# content= "\u5f88\u62b1\u6b49\uff0c\u6211\u65e0\u6cd5\u63d0\u4f9b\u5173\u4e8e\u5728\u676d\u5dde\u7f34\u7a0e\u7684\u8be6\u7ec6\u4fe1\u606f\u3002\u8bf7\u60a8\u54a8\u8be2\u5f53\u5730\u7a0e\u52a1\u5c40\u6216\u8005\u76f8\u5173\u7a0e\u52a1\u90e8\u95e8\uff0c\u4ed6\u4eec\u5c06\u80fd\u591f\u4e3a\u60a8\u63d0\u4f9b\u6b63\u786e\u7684\u4fe1\u606f\u548c\u6307\u5bfc\u3002"
#
# # print(p2.encode().decode("unicode_escape"))
# # print(p2.encode('utf-8').decode('unicode_escape'))
# print(str.encode('unicode-escape').decode('unicode-escape'))
- 测试信息总结
import openai
import json
'''
尝试function函数对input进行总结，方便后续用总结后的内容，发起其他操作。
-- 结果：
可以联系上下文，但是仍然给出的是最后一个问题的回复，而不是内容总结！！！
'''
def get_summary(summary):
tax_info = {
"summary": summary
}
return json.dumps(tax_info)
# Step 1, send model the user query and what functions it has access to
def run_conversation(messages):
response = openai.ChatCompletion.create(
model="gpt-3.5-turbo-0613",
# messages=[{"role": "user", "content": content}],
messages=messages,
functions=[
{
"name": "get_summary",
"description": "提炼输入内容的主要信息，总结成句子的中心思想并输出",
"parameters": {
"type": "object",
"properties": {
"summary": {
"type": "string",
"description": "对输入内容的一段总结性描述，里面包含了输入内容的重要信息，不重要或非相关的信息被丢弃",
}
},
"required": ["summary"],
},
}
],
function_call="auto",
)
message = response["choices"][0]["message"]
print("*" * 60)
print(message)
# print(type(message)) # <class 'openai.openai_object.OpenAIObject'>
print("*" * 60)
arguments_ = message["content"]
# print(type(arguments_)) # <class 'str'>
# print(arguments_) # 这里已经是正常显示的中文了，所以不需要对字符串再次unicode解码
return arguments_
old_q = "我想邀请员工线上签合同，要怎么在你们app上操作"
old_a = "您好，您先在【首页】点击【人员管理】-【员工】，选择线上签署签约，对合同模板内容进行确认并输入验证码获取签章，然后填写合同相关内容信息（被邀请的人员姓名、身份证号、税前工资、发薪日），最后对合同内容确认无误后即可通过链接或二维码方式发起邀请"
new_q = "那如果是线下呢"
# content = "我在杭州上班，也在这里缴税"
messages = [
{"role": "user", "content": old_q},
{"role": "assistant", "content": old_a},
{"role": "user", "content": new_q},
]
result = run_conversation(messages)
print(result)
