Langchain Chat Model 和 Chat Prompt Template

0. 简介

Chat Model 不止是一个用于聊天对话的模型抽象，更重要的是提供了多角色提示能力（System,AI,Human,Function)。
Chat Prompt Template 则为开发者提供了便捷维护不同角色的提示模板与消息记录的接口。

1. 构造 ChatPromptTemplate

from langchain.prompts.chat import (
    ChatPromptTemplate,
    SystemMessagePromptTemplate,
    AIMessagePromptTemplate,
   HumanMessagePromptTemplate,
)

import os
from dotenv import load_dotenv, find_dotenv

# 删除all_proxy环境变量
if 'all_proxy' in os.environ:
    del os.environ['all_proxy']

# 删除ALL_PROXY环境变量
if 'ALL_PROXY' in os.environ:
    del os.environ['ALL_PROXY']

_ = load_dotenv(find_dotenv())

template = (
    """You are a translation expert, proficient in various languages. \n
    Translates English to Chinese."""
)
system_message_prompt = SystemMessagePromptTemplate.from_template(template)
print(type(system_message_prompt))
print(system_message_prompt)

human_template = "{text}"
human_message_prompt = HumanMessagePromptTemplate.from_template(human_template)
print(type(human_message_prompt))
print(human_message_prompt)

print("*"*40)
# 使用 System 和 Human 角色的提示模板构造 ChatPromptTemplate
chat_prompt_template = ChatPromptTemplate.from_messages(
    [system_message_prompt, human_message_prompt]
)
print(type(chat_prompt_template))
print(chat_prompt_template)

print("*"*50)
chat_prompt_prompt_value = chat_prompt_template.format_prompt(text="I love python.")
print(type(chat_prompt_prompt_value))
print(chat_prompt_prompt_value)

print("*"*60)
chat_prompt_list = chat_prompt_template.format_prompt(text="I love python.").to_messages()
print(type(chat_prompt_list))
print(chat_prompt_list)

输出：

<class 'langchain_core.prompts.chat.SystemMessagePromptTemplate'>
prompt=PromptTemplate(input_variables=[], input_types={}, partial_variables={}, template='You are a translation expert, proficient in various languages. \n\n    Translates English to Chinese.') additional_kwargs={}
<class 'langchain_core.prompts.chat.HumanMessagePromptTemplate'>
prompt=PromptTemplate(input_variables=['text'], input_types={}, partial_variables={}, template='{text}') additional_kwargs={}
****************************************
<class 'langchain_core.prompts.chat.ChatPromptTemplate'>
input_variables=['text'] input_types={} partial_variables={} messages=[SystemMessagePromptTemplate(prompt=PromptTemplate(input_variables=[], input_types={}, partial_variables={}, template='You are a translation expert, proficient in various languages. \n\n    Translates English to Chinese.'), additional_kwargs={}), HumanMessagePromptTemplate(prompt=PromptTemplate(input_variables=['text'], input_types={}, partial_variables={}, template='{text}'), additional_kwargs={})]
**************************************************
<class 'langchain_core.prompt_values.ChatPromptValue'>
messages=[SystemMessage(content='You are a translation expert, proficient in various languages. \n\n    Translates English to Chinese.', additional_kwargs={}, response_metadata={}), HumanMessage(content='I love python.', additional_kwargs={}, response_metadata={})]
************************************************************
<class 'list'>
[SystemMessage(content='You are a translation expert, proficient in various languages. \n\n    Translates English to Chinese.', additional_kwargs={}, response_metadata={}), HumanMessage(content='I love python.', additional_kwargs={}, response_metadata={})]

2. LCEL 执行

from langchain_openai import ChatOpenAI
from langchain_core.output_parsers import StrOutputParser
from langchain.prompts.chat import (
    ChatPromptTemplate,
    SystemMessagePromptTemplate,
    AIMessagePromptTemplate,
   HumanMessagePromptTemplate,
)
import os
from dotenv import load_dotenv, find_dotenv

# 删除all_proxy环境变量
if 'all_proxy' in os.environ:
    del os.environ['all_proxy']

# 删除ALL_PROXY环境变量
if 'ALL_PROXY' in os.environ:
    del os.environ['ALL_PROXY']

_ = load_dotenv(find_dotenv())

# 为了结果的稳定性，将 temperature 设置为 0
translation_model = ChatOpenAI(model_name="gpt-3.5-turbo", temperature=0)

template = (
    """You are a translation expert, proficient in various languages. \n
    Translates {source_language} to {target_language} in the style of {name}."""
)
system_message_prompt = SystemMessagePromptTemplate.from_template(template)

human_template = "{text}"
human_message_prompt = HumanMessagePromptTemplate.from_template(human_template)

# 使用 System 和 Human 角色的提示模板构造 ChatPromptTemplate
m_chat_prompt_template = ChatPromptTemplate.from_messages(
    [system_message_prompt, human_message_prompt]
)
output_parser = StrOutputParser()

m_translation_chain = m_chat_prompt_template| translation_model |  StrOutputParser(callbacks=[callback_handler])

# Prepare input data
input_data = {
    "source_language": "English",
    "target_language": "Chinese",
    "name": "严复",
    "text": "Life is full of regrets. All we can do is to minimize them.",
}
input_data1 = {
    "source_language": "English",
    "target_language": "Chinese",
    "name": "李白",
    "text": "Life is full of regrets. All we can do is to minimize them.",
}


# Format the prompt
prompt_value = m_chat_prompt_template.format_prompt(**input_data)
print(type(prompt_value))
print(prompt_value)

print(type(prompt_value.to_messages()))
print(prompt_value.to_messages())

result = translation_model.invoke(prompt_value)
print(result)
result = m_translation_chain.invoke(input_data)
print(result)
result = m_translation_chain.invoke(input_data1)
print(result)

输出：

<class 'langchain_core.prompt_values.ChatPromptValue'>
messages=[SystemMessage(content='You are a translation expert, proficient in various languages. \n\n    Translates English to Chinese in the style of 严复.', additional_kwargs={}, response_metadata={}), HumanMessage(content='Life is full of regrets. All we can do is to minimize them.', additional_kwargs={}, response_metadata={})]
<class 'list'>
[SystemMessage(content='You are a translation expert, proficient in various languages. \n\n    Translates English to Chinese in the style of 严复.', additional_kwargs={}, response_metadata={}), HumanMessage(content='Life is full of regrets. All we can do is to minimize them.', additional_kwargs={}, response_metadata={})]
content='人生充满了遗憾。我们所能做的就是尽量减少它们。' additional_kwargs={'refusal': None} response_metadata={'token_usage': {'completion_tokens': 31, 'prompt_tokens': 53, 'total_tokens': 84, 'completion_tokens_details': None, 'prompt_tokens_details': None}, 'model_name': 'gpt-3.5-turbo-0125', 'system_fingerprint': None, 'finish_reason': 'stop', 'logprobs': None} id='run-676a9818-bbbc-44f7-a30e-0ec065aa502f-0' usage_metadata={'input_tokens': 53, 'output_tokens': 31, 'total_tokens': 84, 'input_token_details': {}, 'output_token_details': {}}
人生充满了遗憾。我们所能做的就是尽量减少它们。
人生充滿遺憾，唯有盡量減少。

当 translation_model.invoke(input_data) 输入参数是字典时报错

result = translation_model.invoke(input_data)
print(result)

报错：

ValueError: Invalid input type <class 'dict'>. Must be a PromptValue, str, or list of BaseMessages.

当 m_translation_chain.invoke(prompt_value) 输入参数是prompt_value时报错

result = m_translation_chain.invoke(prompt_value)
print(result)

报错：

TypeError: Expected mapping type as input to ChatPromptTemplate. Received <class 'langchain_core.prompt_values.ChatPromptValue'>.

查看属性：

input_schema = m_translation_chain.input_schema.model_json_schema()
print(input_schema)
output_schema = m_translation_chain.output_schema.model_json_schema()
print(output_schema)

输出：

{'properties': {'name': {'title': 'Name', 'type': 'string'}, 'source_language': {'title': 'Source Language', 'type': 'string'}, 'target_language': {'title': 'Target Language', 'type': 'string'}, 'text': {'title': 'Text', 'type': 'string'}}, 'required': ['name', 'source_language', 'target_language', 'text'], 'title': 'PromptInput', 'type': 'object'}

{'title': 'StrOutputParserOutput', 'type': 'string'}