备忘，LangChain建立本地知识库的几个要点

news2024/11/28 6:54:00

本地知识库可以解决本地资源与AI结合的问题，为下一步应用管理已有资产奠定基础。本地知识库的建立可参考LangChain结合通义千问的自建知识库（二）、（三）、（四）
本文主要记录两个方面的问题
1 搭建过程中遇到的坑
2 向量是数据库改成ES7

1 搭建过程中遇到的坑
1）安装bce-embedding-base_v1模型
需要用git clone到本地，但由于模型比较大，需要先安装git lfs管理大型的文件，再克隆
sudo apt-get install git-lfs

安装成功后，再进入预备安装模型的目录下，执行clone
git clone https://www.modelscope.cn/maidalun/bce-embedding-base_v1.git

nltk corpora语料库缺失
报错样例如下：

缺少两个库 punkt和 averaged_perceptron_tagger
解决的方式是离线下载，参考NLTK：离线安装punkt

2 向量是数据库改成ES7
原博中使用Chroma，但我已经装了ES7，所以改用ES7作为向量数据库
原博中两部分代码，导入本地知识库和应用本地知识库，修改如下

导入本地知识库

import time
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain_community.document_loaders import UnstructuredFileLoader
from langchain_community.embeddings.huggingface import HuggingFaceEmbeddings
from langchain_community.vectorstores import ElasticVectorSearch

time_list = []
t = time.time()

# loader = UnstructuredFileLoader("/home/cfets/AI/textwarefare/test.txt")
loader = UnstructuredFileLoader("/home/cfets/AI/textwarefare/cmtest.txt")
data = loader.load()

# 文本切分去。
text_splitter = RecursiveCharacterTextSplitter(chunk_size=50, chunk_overlap=0)
split_docs = text_splitter.split_documents(data)
print(split_docs)

model_name = r"/home/cfets/AI/model/bce-embedding-base_v1"
model_kwargs = {'device': 'cpu'}
encode_kwargs = {'normalize_embeddings': False}

embeddings = HuggingFaceEmbeddings(
    model_name=model_name,
    model_kwargs=model_kwargs,
    encode_kwargs=encode_kwargs)

# 初始化加载器 构建本地知识向量库
# db = Chroma.from_documents(split_docs, embeddings, persist_directory="./chroma/bu_test_bec")

# 使用ES
db = ElasticVectorSearch.from_documents(
    split_docs,
    embeddings,
    elasticsearch_url="http://localhost:9200",
    index_name="elastic-ai-test",
)

# 持久化
#  db.persist()
print(db.client.info())

# 打印时间##
time_list.append(time.time() - t)
print(time.time() - t)

应用本地知识库

import dashscope
from dashscope import Generation
from dashscope.api_entities.dashscope_response import Role
from http import HTTPStatus
from langchain_community.vectorstores import Chroma
from langchain_community.vectorstores import ElasticVectorSearch
from langchain_community.embeddings.huggingface import HuggingFaceEmbeddings

dashscope.api_key = XXXXXX


def conversation_mutual():
    messages = []

    # 引入模型
    model_name = r"/home/cfets/AI/model/bce-embedding-base_v1"
    model_kwargs = {'device': 'cpu'}
    encode_kwargs = {'normalize_embeddings': False}

    embeddings = HuggingFaceEmbeddings(
        model_name=model_name,
        model_kwargs=model_kwargs,
        encode_kwargs=encode_kwargs
    )

    # 使用chroma本地库
    # db = Chroma(persist_directory="./chroma/bu_cmtest_bec", embedding_function=embeddings)

    # 使用elastic search
    my_index = "es_akshare-api"
    db = ElasticVectorSearch(
        embedding=embeddings,
        elasticsearch_url="http://localhost:9200",
        index_name=my_index,
    )


    while True:
        message = input('user:')

        # 引入本地库
        similarDocs = db.similarity_search(message, k=5)
        summary_prompt = "".join([doc.page_content for doc in similarDocs])
        print(summary_prompt)
        send_message = f"下面的信息({summary_prompt})是否有这个问题({message})有关" \
                       f"，如果你觉得无关请告诉我无法根据提供的上下文回答'{message}'这个问题，简要回答即可" \
                       f"，否则请根据{summary_prompt}对{message}的问题进行回答"

        messages.append({'role': Role.USER, 'content': message})
        messages.append({'role': Role.USER, 'content': send_message})  # 按本地库回复

        whole_message = ''
        # 切换到通义模型
        responses = Generation.call(Generation.Models.qwen_plus, messages=messages, result_format='message', stream=True,
                                    incremental_output=True)
        print('system:', end='')
        for response in responses:
            if response.status_code == HTTPStatus.OK:
                whole_message += response.output.choices[0]['message']['content']
                print(response.output.choices[0]['message']['content'], end='')
            else:
                print('Request id: %s, Status code: %s, error code: %s, error message: %s' % (
                    response.request_id, response.status_code,
                    response.code, response.message
                ))

    print()
    messages.append({'role': 'assistant', 'content': whole_message})
    print('\n')


if __name__ == '__main__':
    conversation_mutual()