探索MariTalk与LangChain的结合：从基础到进阶

afTFODguAKBF

于 2024-10-06 11:58:20 发布

阅读量98

点赞数 1

文章标签： langchain python

本文链接：https://blog.csdn.net/afTFODguAKBF/article/details/142725320

版权

探索MariTalk与LangChain的结合：从基础到进阶

引言

在这篇文章中，我们将探讨如何利用巴西公司Maritaca AI开发的语言助手MariTalk与LangChain的结合，实现多种任务自动化。我们将从简单的宠物命名任务开始，深入到复杂的文本生成和问答系统——特别是在处理超过MariTalk令牌限制的长文档时。

主要内容

安装

首先，我们需要安装LangChain库和其依赖项：

!pip install langchain langchain-core langchain-community httpx

获取API密钥

您需要访问chat.maritaca.ai获取API密钥。

示例1 - 宠物名字建议

定义我们的语言模型 ChatMaritalk 并配置API密钥：

from langchain_community.chat_models import ChatMaritalk
from langchain_core.output_parsers import StrOutputParser
from langchain_core.prompts.chat import ChatPromptTemplate

llm = ChatMaritalk(
    model="sabia-2-medium",
    api_key="your_api_key",  # 在这里插入您的API密钥
    temperature=0.7,
    max_tokens=100,
)

output_parser = StrOutputParser()

chat_prompt = ChatPromptTemplate.from_messages(
    [
        (
            "system",
            "You are an assistant specialized in suggesting pet names. Given the animal, you must suggest 4 names.",
        ),
        ("human", "I have a {animal}"),
    ]
)

chain = chat_prompt | llm | output_parser

response = chain.invoke({"animal": "dog"})
print(response)  # 输出类似 "1. Max\n2. Bella\n3. Charlie\n4. Rocky"

流式生成

对于长文本生成任务，可以采用同步和异步方式接收响应：

同步：

from langchain_core.messages import HumanMessage

messages = [HumanMessage(content="Suggest 3 names for my dog")]

for chunk in llm.stream(messages):
    print(chunk.content, end="", flush=True)

异步：

from langchain_core.messages import HumanMessage
import asyncio

async def async_invoke_chain(animal: str):
    messages = [HumanMessage(content=f"Suggest 3 names for my {animal}")]
    async for chunk in llm._astream(messages):
        print(chunk.message.content, end="", flush=True)

await async_invoke_chain("dog")

示例2 - RAG + LLM：UNICAMP 2024考试问答系统

加载数据库

我们首先创建一个数据库，并将下载的PDF文本分割成500字符的块。

from langchain_community.document_loaders import OnlinePDFLoader
from langchain_text_splitters import RecursiveCharacterTextSplitter

loader = OnlinePDFLoader(
    "https://www.comvest.unicamp.br/wp-content/uploads/2023/10/31-2023-Dispoe-sobre-o-Vestibular-Unicamp-2024_com-retificacao.pdf"
)
data = loader.load()

text_splitter = RecursiveCharacterTextSplitter(
    chunk_size=500, chunk_overlap=100, separators=["\n", " ", ""]
)
texts = text_splitter.split_documents(data)

创建搜索器

使用BM25作为简易搜索系统：

from langchain_community.retrievers import BM25Retriever

retriever = BM25Retriever.from_documents(texts)

结合搜索系统和LLM

实现提示并调用链：

from langchain.chains.question_answering import load_qa_chain

prompt = """Baseado nos seguintes documentos, responda a pergunta abaixo.

{context}

Pergunta: {query}
"""

qa_prompt = ChatPromptTemplate.from_messages([("human", prompt)])

chain = load_qa_chain(llm, chain_type="stuff", verbose=True, prompt=qa_prompt)

query = "Qual o tempo máximo para realização da prova?"
docs = retriever.invoke(query)

chain.invoke(
    {"input_documents": docs, "query": query}
)