langchain结合searXNG实现基于搜索RAG

目录

一、背景

二、环境说明和安装

1.1 环境说明

2.2 环境安装

2.2.1 searXNG安装

三、代码实现

代码

结果输出

直接请求模型输出

四、参考

---

一、背景

        大语言模型的出现带来了新的技术革新，但是大模型由于训练语料的原因，它的知识和当前实时热点存在时间差距，存在很严重的幻觉。而RAG检索增强生成能够解决这个问题，知识以prompt的形式进行补足，再让LLM进行润色进而呈现给用户。

二、环境说明和安装

1.1 环境说明

        整个环境都是基于我本地进行搭建，机器环境采用的是wsl2部署的Ubuntu-20.04，大模型选取的是Qwen/Qwen2-7B-Instruct，采用vllm进行部署推理以openai api的形式提供服务，langchain使用的是最新的0.2版本，由于searXNG是需要docker部署的，所以请确保已安装docker .desktop。

2.2 环境安装

2.2.1 searXNG安装

• 下载

cd /opt/softwares
git clone https://github.com/searxng/searxng-docker.git
cd searxng-docker

• docker-compose配置

searXNG的组件有三个，caddy实现反向代理，这里不需要反向代理，采用直接调用的形式，关于caddy的部分注释掉；redis用来保存session，searXNG提供搜索服务。设置端口映射，0.0.0.0:8080:8080左边的端口为宿主机端口，右边的端口为docker端口

version: "3.7"

services:
        #caddy:
        #container_name: caddy
        #image: docker.io/library/caddy:2-alpine
        #network_mode: host
        #restart: unless-stopped
        #volumes:
        #- ./Caddyfile:/etc/caddy/Caddyfile:ro
        #- caddy-data:/data:rw
        #- caddy-config:/config:rw
        #environment:
        #- SEARXNG_HOSTNAME=${SEARXNG_HOSTNAME:-http://localhost:80}
        #- SEARXNG_TLS=${LETSENCRYPT_EMAIL:-internal}
        #cap_drop:
        #- ALL
        #cap_add:
        #- NET_BIND_SERVICE
        #logging:
        #driver: "json-file"
        #options:
        #max-size: "1m"
        #max-file: "1"

  redis:
    container_name: redis
    image: docker.io/valkey/valkey:7-alpine
    command: valkey-server --save 30 1 --loglevel warning
    restart: unless-stopped
    networks:
      - searxng
    volumes:
      - valkey-data2:/data
    cap_drop:
      - ALL
    cap_add:
      - SETGID
      - SETUID
      - DAC_OVERRIDE
    logging:
      driver: "json-file"
      options:
        max-size: "1m"
        max-file: "1"

  searxng:
    container_name: searxng
    image: docker.io/searxng/searxng:latest
    restart: unless-stopped
    networks:
      - searxng
    ports:
      - "0.0.0.0:8080:8080"
    volumes:
      - ./searxng:/etc/searxng:rw
    environment:
      - SEARXNG_BASE_URL=https://${SEARXNG_HOSTNAME:-localhost}/
    cap_drop:
      - ALL
    cap_add:
      - CHOWN
      - SETGID
      - SETUID
    logging:
      driver: "json-file"
      options:
        max-size: "1m"
        max-file: "1"

networks:
  searxng:

volumes:
  caddy-data:
  caddy-config:
  valkey-data2:

• searxng配置

# 设置秘钥
sed -i "s|ultrasecretkey|$(openssl rand -hex 32)|g" searxng/settings.yml
# limiter设置为false，设置formats
vim searxng/settings.yml

settings.yml内容如下：

use_default_settings: true
server:
  # base_url is defined in the SEARXNG_BASE_URL environment variable, see .env and docker-compose.yml
  secret_key: "fc6a61736b885ace5f990f840006d057c01604501a36636b4a378f4b2cb89fdf"  # change this!
  limiter: false  # can be disabled for a private instance
  image_proxy: true
ui:
  static_use_hash: true
redis:
  url: redis://redis:6379/0
search:
  formats:
    - html
    - json

如果没有配置search.formats，在使用langchain进行调用的时候会报503错误，相关issue如下：

https://github.com/langchain-ai/langchain/issues/855

• 启动

# 启动
docker-compose up -d
# 停止
docker-compose stop

• 访问localhost:8080，如下：

三、代码实现

代码

from typing import List
from langchain_community.utilities import SearxSearchWrapper
import requests
from requests.exceptions import RequestException
from langchain_community.document_loaders import AsyncHtmlLoader
from langchain_community.document_transformers.html2text import Html2TextTransformer
from langchain_text_splitters import RecursiveCharacterTextSplitter
from langchain_openai.chat_models import ChatOpenAI
from langchain.prompts.chat import ChatPromptTemplate
import faiss
from langchain_community.vectorstores import FAISS
from langchain_community.docstore import InMemoryDocstore
from langchain_huggingface import HuggingFaceEmbeddings
from langchain_core.documents import Document
from langchain.chains.combine_documents import create_stuff_documents_chain
from langchain.chains.retrieval import create_retrieval_chain
import os
import sys
sys.path.append(os.path.abspath(os.pardir))
from global_config import MODEL_CACHE_DIR
os.environ["KMP_DUPLICATE_LIB_OK"] = "TRUE"


def check_url_access(urls: List[str]) -> List[str]:
    urls_can_access = []
    for url in urls:
        try:
            req_res = requests.get(url)
            if req_res is not None and req_res.ok:
                urls_can_access.append(url)
        except RequestException as e:
            continue
    return urls_can_access


def get_chat_llm(streaming=False, temperature=0, max_tokens=2048) -> ChatOpenAI:
    chat_llm = ChatOpenAI(
        model_name="Qwen2-7B-Instruct",
        openai_api_key="empty",
        openai_api_base="http://localhost:8000/v1",
        max_tokens=max_tokens,
        temperature=temperature,
        streaming=streaming
    )
    return chat_llm


def get_retriever(documents: List[Document]):
    # embedding
    embeddings = HuggingFaceEmbeddings(
        model_name=os.path.join(MODEL_CACHE_DIR, "maidalun/bce-embedding-base_v1")
    )
    # faiss
    faiss_index = faiss.IndexFlatL2(len(embeddings.embed_query("hello world")))  # 指定向量维度
    vector_store = FAISS(
        embedding_function=embeddings,
        index=faiss_index,
        docstore=InMemoryDocstore(),
        index_to_docstore_id={}
    )
    vector_store.add_documents(documents)
    return vector_store.as_retriever()


def get_answer_prompt() -> ChatPromptTemplate:
    system_prompt = """
    你是一个问答任务的助手，请依据以下检索出来的信息去回答问题：
    {context}
    """
    return ChatPromptTemplate([
        ("system", system_prompt),
        ("human", "{input}")
    ])


if __name__ == '__main__':
    llm = get_chat_llm()
    search = SearxSearchWrapper(searx_host="http://localhost:8080/")
    query = "《元梦之星》这款游戏是什么时候发行的"
    results = search.results(
        query,                                  # 搜索内容
        language="zh-CN",                       # 语言
        safesearch=1,                           # 可选0(不过滤任何结果),1(中等级别内容过滤),2(严格级别内容过滤)
        categories="general",                   # 搜索内容，取值general/images/videos等
        engines=["bing", "brave", "google"],    # 搜索引擎
        num_results=3                           # 返回内容数
    )
    print(f"search results: {results}")

    urls_to_look = [ele["link"] for ele in results if ele.get("link", None)]
    # 检查url是否可达
    urls = check_url_access(urls_to_look)
    print(f"urls: {urls}")
    # loader
    loader = AsyncHtmlLoader(urls, ignore_load_errors=True, requests_kwargs={"timeout": 5})
    docs = loader.load()
    # transformer
    html2text = Html2TextTransformer()
    docs = html2text.transform_documents(docs)

    # text split
    text_splitter = RecursiveCharacterTextSplitter(chunk_size=500, chunk_overlap=100)
    docs = text_splitter.split_documents(docs)

    # retriever
    retriever = get_retriever(docs)

    # create_stuff_documents_chain的prompt必须含有context变量
    qa_chain = create_stuff_documents_chain(llm, get_answer_prompt())
    rag_chain = create_retrieval_chain(retriever, qa_chain)

    res = rag_chain.invoke({
        "input": query
    })
    print(res["answer"])

结果输出

《元梦之星》这款游戏是在2023年12月15日全平台上线发行的。

直接请求模型输出

curl http://localhost:8000/v1/chat/completions \
    -H "Content-Type: application/json" \
    -d '{
    "model": "Qwen2-7B-Instruct",
    "messages": [
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "《元梦之星》这款游戏是什么时候发行的"}
    ]
    }'

从基于搜索的RAG输出和模型原始输出来看，搜索能补充大模型缺失的知识，从而基于新知识来进行回答

四、参考

https://github.com/searxng/searxng-docker

https://github.com/ptonlix/LangChain-SearXNG

SearxNG Search | 🦜️🔗 LangChain

Conversational RAG | 🦜️🔗 LangChain

maidalun1020/bce-embedding-base_v1 · Hugging Face