LangChain实现聊天机器人

little_pierce

于 2024-06-30 18:49:22 发布

阅读量1.2k

点赞数 13

分类专栏： Agent 文章标签： langchain prompt

本文链接：https://blog.youkuaiyun.com/qq_43668800/article/details/140084487

版权

Agent 专栏收录该内容

1 篇文章

订阅专栏

项目的接口模型是GPT-3.5

1 接入大模型

from langchain_openai import ChatOpenAI

llm = ChatOpenAI(
    openai_api_base="xxxxxxxxxxxxxxxx",
    openai_api_key="sk-xxxxxxxxxxxxxxxxxxxxxxxxxxx",
)

可以看到，模型对打招呼做出返回：

from langchain_core.messages import HumanMessage

llm.invoke([HumanMessage(content="Hi! I'm Bob")])

AIMessage(content=‘Hello, Bob! Nice to meet you. How can I assist you today?’, response_metadata={‘token_usage’: {‘completion_tokens’: 0, ‘prompt_tokens’: 0, ‘total_tokens’: 0}, ‘model_name’: ‘gpt-3.5-turbo’, ‘system_fingerprint’: None, ‘finish_reason’: ‘stop’, ‘logprobs’: None}, id=‘run-3f0b67d0-0900-427c-9eda-9e405893be60-0’)

但是此时再问人类的姓名，AI是不知道的，因为默认没有记忆功能

llm.invoke([HumanMessage(content="What's my name?")])

AIMessage(content=“I’m sorry, I don’t have access to personal information like your name unless you tell me. How can I assist you today?”, response_metadata={‘token_usage’: {‘completion_tokens’: 0, ‘prompt_tokens’: 0, ‘total_tokens’: 0}, ‘model_name’: ‘gpt-3.5-turbo’, ‘system_fingerprint’: None, ‘finish_reason’: ‘stop’, ‘logprobs’: None}, id=‘run-0a75ae65-4202-4a69-bd67-859aa4be676c-0’)

此处还要注意另一点，invoke方法这里调用的参数是input，其接受类型为Union[PromptValue, str, Sequence[MessageLikeRepresentation]]，也就是说PromptValue、字符串、消息列表都可以。上面实际上接受的是消息列表，下面尝试接受另两个类型：

2 多种输入类型作为prompt调用大模型

llm.invoke("Hello? Who are you?") # 此处是字符串类型调用

AIMessage(content=“Hello! I’m ChatGPT, an AI language model. I’m here to help answer questions, have conversations, or assist with whatever you need. How can I help you today?”, response_metadata={‘token_usage’: {‘completion_tokens’: 0, ‘prompt_tokens’: 0, ‘total_tokens’: 0}, ‘model_name’: ‘gpt-3.5-turbo’, ‘system_fingerprint’: None, ‘finish_reason’: ‘stop’, ‘logprobs’: None}, id=‘run-035271f6-6633-493e-ba27-ec1707c188b2-0’)

from langchain_core.prompts import PromptTemplate

# 此处是PromptValue类型调用，其中PromptValue本身是抽象类，StringPromptValue是字符串类型prompt的对其实现
str = "Can you translate {content} to Chinese?"
promptTemplate = PromptTemplate.from_template(str)
prompt = promptTemplate.format(content = "pierce")
print(promptTemplate, prompt)
llm.invoke(prompt)

input_variables=[‘content’] template=‘Can you translate {content} to Chinese?’ Can you translate pierce to Chinese?
AIMessage(content=‘The word “pierce” can be translated to Chinese as 刺穿 (cì chuān) or 穿透 (chuān tòu), depending on the context in which it’s used.’, response_metadata={‘token_usage’: {‘completion_tokens’: 0, ‘prompt_tokens’: 0, ‘total_tokens’: 0}, ‘model_name’: ‘gpt-3.5-turbo’, ‘system_fingerprint’: None, ‘finish_reason’: ‘stop’, ‘logprobs’: None}, id=‘run-fb10b27a-4491-443f-a233-ea48910ad089-0’)

回到消息历史方面，我们可以将历史对话传给模型，让其看起来“有记忆功能”：

3 为其添加记忆功能

我们以消息列表的形式将历史记录传递给模型，从而实现记忆功能：

from langchain_core.messages import AIMessage

llm.invoke(
    [
        HumanMessage(content="Hi! I'm Pierce"),
        AIMessage(content="Hello Pierce! How can I assist you today?"),
        HumanMessage(content="What's my name?"),
    ]
)

AIMessage(content=‘Your name is Pierce!’, response_metadata={‘token_usage’: {‘completion_tokens’: 0, ‘prompt_tokens’: 0, ‘total_tokens’: 0}, ‘model_name’: ‘gpt-3.5-turbo’, ‘system_fingerprint’: None, ‘finish_reason’: ‘stop’, ‘logprobs’: None}, id=‘run-286ba0d3-a673-4cb1-9522-a03638c97c0c-0’)

这样显示地将对话历史作为消息列表prompt传给模型有点繁琐，合适的方法是这样的：

from langchain_core.chat_history import InMemoryChatMessageHistory
from langchain_core.runnables.history import RunnableWithMessageHistory

# store 存储所有对话历史，作为memory
store = {}


def get_session_history(session_id: str) -> InMemoryChatMessageHistory:
    if session_id not in store:
        store[session_id] = InMemoryChatMessageHistory()
    return store[session_id]


with_message_history = RunnableWithMessageHistory(llm, get_session_history)

这里，BaseChatMessageHistory作为基类是一个抽象类，要实现任何一个对话历史存储，都要继承自该抽象类，并实现下列方法中的一个或多个：

add_messages: sync variant for bulk addition of messages
aadd_messages: async variant for bulk addition of messages
messages: sync variant for getting messages
aget_messages: async variant for getting messages
clear: sync variant for clearing messages
aclear: async variant for clearing messages

InMemoryChatMessageHistory()方法继承自抽象类并实现了上述方法

这里get_session_history()方法确定如何获取需要的历史信息，session_id的设置用于区分不同的对话信息对于RunnableWithMessageHistory，其本质是一个Runnable，用于管理另一个Runnable的聊天消息历史记录。

RunnableWithMessageHistory包装另一个Runnable并为其管理聊天消息历史；它负责读取和更新聊天消息历史记录

调用时，必须始终使用包含聊天消息历史工厂的适当参数的配置来调用RunnableWithMessageHistory

默认情况下，Runnable需要一个名为“session_id”的配置参数，该参数是一个字符串。该参数用于创建新的或查找与给定session_id匹配的现有聊天消息历史记录

config = {"configurable": {"session_id": "test001"}}

response = with_message_history.invoke(
    [HumanMessage(content="Hi! I'm Pierce")],
    config=config,
)

response.content

Parent run 7c9e16c1-5f7f-44ef-97fc-98c97d2a4de2 not found for run e67007cd-0748-420b-8b16-44641b397ed8. Treating as a root run.
‘Hello Pierce! How can I assist you today?’

response = with_message_history.invoke(
    [HumanMessage(content="What's my name?")],
    config=config,
)

response.content

Parent run 4cb949d7-fc74-40f7-b7cf-a409637f761e not found for run bf0dba2c-930d-4166-b22b-5bd068fb8998. Treating as a root run.
‘Your name is Pierce.’

可以看到，现在模型具有了真正的记忆能力。如果我们修改session_id，即置换到另一个对话的记忆，模型对该次对话的记忆就消失：

config02 = {"configurable": {"session_id": "test002"}}

response = with_message_history.invoke(
    [HumanMessage(content="What's my name?")],
    config=config02,
)

response.content

Parent run f6da62fa-3db5-4e11-a7a4-819e87b9fc15 not found for run be939cc3-29f3-4d9a-bdb6-21bed23cbd4d. Treating as a root run.
“I’m sorry, but I do not have access to personal information about individuals unless it has been shared with me in the course of our conversation. I am designed to respect user privacy and confidentiality.”

只要我们切换回来，就又可以实现记忆了：

config = {"configurable": {"session_id": "test001"}}

response = with_message_history.invoke(
    [HumanMessage(content="What's my name?")],
    config=config,
)

response.content

Parent run beb8df26-8f44-4e21-8982-9f321993caec not found for run 59eb38b9-d329-44eb-9a4e-232e7ff8b997. Treating as a root run.
‘Your name is Pierce.’

因此，将记忆存储在磁盘内做持久化处理也就是可行的了。让我们看看store里存储的信息：

{‘test001’: InMemoryChatMessageHistory(messages=[HumanMessage(content=“Hi! I’m Pierce”), AIMessage(content=‘Hello Pierce! How can I assist you today?’, response_metadata={‘token_usage’: {‘completion_tokens’: 10, ‘prompt_tokens’: 12, ‘total_tokens’: 22}, ‘model_name’: ‘gpt-3.5-turbo’, ‘system_fingerprint’: None, ‘finish_reason’: ‘stop’, ‘logprobs’: None}, id=‘run-e67007cd-0748-420b-8b16-44641b397ed8-0’), HumanMessage(content=“What’s my name?”), AIMessage(content=‘Your name is Pierce.’, response_metadata={‘token_usage’: {‘completion_tokens’: 5, ‘prompt_tokens’: 35, ‘total_tokens’: 40}, ‘model_name’: ‘gpt-3.5-turbo’, ‘system_fingerprint’: None, ‘finish_reason’: ‘stop’, ‘logprobs’: None}, id=‘run-bf0dba2c-930d-4166-b22b-5bd068fb8998-0’), HumanMessage(content=“What’s my name?”), AIMessage(content=‘Your name is Pierce.’, response_metadata={‘token_usage’: {‘completion_tokens’: 5, ‘prompt_tokens’: 53, ‘total_tokens’: 58}, ‘model_name’: ‘gpt-3.5-turbo’, ‘system_fingerprint’: None, ‘finish_reason’: ‘stop’, ‘logprobs’: None}, id=‘run-59eb38b9-d329-44eb-9a4e-232e7ff8b997-0’)]),
‘test002’: InMemoryChatMessageHistory(messages=[HumanMessage(content=“What’s my name?”), AIMessage(content=“I’m sorry, but I do not have access to personal information about individuals unless it has been shared with me in the course of our conversation. I am designed to respect user privacy and confidentiality.”, response_metadata={‘token_usage’: {‘completion_tokens’: 39, ‘prompt_tokens’: 12, ‘total_tokens’: 51}, ‘model_name’: ‘gpt-3.5-turbo’, ‘system_fingerprint’: None, ‘finish_reason’: ‘stop’, ‘logprobs’: None}, id=‘run-be939cc3-29f3-4d9a-bdb6-21bed23cbd4d-0’)])}

应该注意到，我们每次返回的AIMessage中现实的prompt_tokens都在增加，可见是将对话历史作为prompt的一部分传递给模型的。
那么会遇到另一个问题，如果一直保持着对话历史的全部内容，野蛮增长下很快对话历史就会占满模型的全部token。因此，我们需要有能力对对话历史做出掌控，例如根据时间顺序截取、根据对话提到的频率截取、对对话历史做总结等

4 进一步操控对话历史

我们模拟一个上述场景，将token数量限制为65，因而历史信息显示不全：

from langchain_core.messages import SystemMessage, trim_messages, AIMessage


trimmer = trim_messages(
    max_tokens=65,
    strategy="last",
    token_counter=llm,
    include_system=True,
    allow_partial=False,
    start_on="human",
)

messages = [
    SystemMessage(content="you're a good assistant"),
    HumanMessage(content="hi! I'm bob"),
    AIMessage(content="hi!"),
    HumanMessage(content="I like vanilla ice cream"),
    AIMessage(content="nice"),
    HumanMessage(content="whats 2 + 2"),
    AIMessage(content="4"),
    HumanMessage(content="thanks"),
    AIMessage(content="no problem!"),
    HumanMessage(content="having fun?"),
    AIMessage(content="yes!"),
]

trimmer.invoke(messages)

[SystemMessage(content=“you’re a good assistant”),
HumanMessage(content=‘whats 2 + 2’),
AIMessage(content=‘4’),
HumanMessage(content=‘thanks’),
AIMessage(content=‘no problem!’),
HumanMessage(content=‘having fun?’),
AIMessage(content=‘yes!’)]

这里就是我们之前提到的按时间顺序截取了，其中设置了策略为“last”；此外，设置了开头的系统信息不被丢弃，且保存的对话总是以人类信息开始。这时我们分别问2+2之前的问题和之后的问题，可看到区别：

from operator import itemgetter
from langchain_core.prompts import ChatPromptTemplate, MessagesPlaceholder
from langchain_core.runnables import RunnablePassthrough

prompt = ChatPromptTemplate.from_messages(
    [
        (
            "system",
            "You are a helpful assistant. Answer all questions to the best of your ability in {language}.",
        ),
        MessagesPlaceholder(variable_name="messages"),
    ]
)

chain = (
    RunnablePassthrough.assign(messages=itemgetter("messages") | trimmer)
    | prompt
    | llm
)

response = chain.invoke(
    {
        "messages": messages + [HumanMessage(content="what's my name?")],
        "language": "English",
    }
)
response.content

“I’m sorry, but I don’t have access to personal information, so I don’t know your name.”

response = chain.invoke(
    {
        "messages": messages + [HumanMessage(content="what math problem did i ask")],
        "language": "English",
    }
)
response.content

‘You asked what is 2 + 2.’

可以看到，之前抛弃掉的信息模型不能再看到，但尚未抛弃的是可以看到的

with_message_history = RunnableWithMessageHistory(
    chain,
    get_session_history,
    input_messages_key="messages",
)

config = {"configurable": {"session_id": "newHistory"}}

response = with_message_history.invoke(
    {
        "messages": messages + [HumanMessage(content="whats my name?")],
        "language": "English",
    },
    config=config,
)

response.content

Parent run 87a69d25-9527-4b21-abf0-03f319245768 not found for run 96ba26e2-7c6c-4906-88da-84ecfb3b1665. Treating as a root run.
“I’m sorry, but I don’t have access to personal information.”

我们当然知道模型已经无法得知我们的名字这一事实了，但不要忘记询问名字这一轮对话现在也加入了内存，顶掉了最前面关于2+2的对话记忆！

response = with_message_history.invoke(
    {
        "messages": [HumanMessage(content="what math problem did i ask?")],
        "language": "English",
    },
    config=config,
)

response.content

Parent run 88f791be-b831-4849-b951-d41d23e573eb not found for run 357a5dcd-17be-44bf-aa73-39e94e414db7. Treating as a root run.
“I’m sorry, but I don’t have access to that information. Please let me know what math problem you would like assistance with and I’ll do my best to help you.”

5 流式输出

最后，模拟网页版ChatGPT回答问题时的流式输出样式：

config = {"configurable": {"session_id": "streaming"}}
for r in with_message_history.stream(
    {
        "messages": [HumanMessage(content="hi! I'm todd. tell me a joke")],
        "language": "English",
    },
    config=config,
):
    print(r.content, end="|")

Parent run b95fdd59-d985-4944-bdec-ba12f5a76005 not found for run 5c7c9ba4-3de8-450d-8d2d-0f8c3e7b5068. Treating as a root run.
|Hi| Todd|!| Sure|,| here|'s| a| joke| for| you|:
|Why| don|'t| skeletons| fight| each| other|?
|They| don|'t| have| the| guts|!||

所有chain都暴露了.stream方法，而使用消息历史记录的chain也不例外。可以简单地使用该方法来获取流式响应。