如何手撸一个自有知识库的RAG系统

RAG通常指的是"Retrieval-Augmented Generation"，即“检索增强的生成”。这是一种结合了检索（Retrieval）和生成（Generation）的机器学习模型，通常用于自然语言处理任务，如文本生成、问答系统等。

我们通过一下几个步骤来完成一个基于京东云官网文档的RAG系统

• 数据收集
• 建立知识库
• 向量检索
• 提示词与模型

数据收集

数据的收集再整个RAG实施过程中无疑是最耗人工的，涉及到收集、清洗、格式化、切分等过程。这里我们使用京东云的官方文档作为知识库的基础。文档格式大概这样：

1{
2    "content": "DDoS IP高防结合Web应用防火墙方案说明\n=======================\n\n\nDDoS IP高防+Web应用防火墙提供三层到七层安全防护体系，应用场景包括游戏、金融、电商、互联网、政企等京东云内和云外的各类型用户。\n\n\n部署架构\n====\n\n\n[![\"部署架构\"](\"https://jdcloud-portal.oss.cn-north-1.jcloudcs.com/cn/image/Advanced%20Anti-DDoS/Best-Practice02.png\")](\"https://jdcloud-portal.oss.cn-north-1.jcloudcs.com/cn/image/Advanced%20Anti-DDoS/Best-Practice02.png\")  \n\nDDoS IP高防+Web应用防火墙的最佳部署架构如下：\n\n\n* 京东云的安全调度中心，通过DNS解析，将用户域名解析到DDoS IP高防CNAME。\n* 用户正常访问流量和DDoS攻击流量经过DDoS IP高防清洗，回源至Web应用防火墙。\n* 攻击者恶意请求被Web应用防火墙过滤后返回用户源站。\n* Web应用防火墙可以保护任何公网的服务器，包括但不限于京东云，其他厂商的云，IDC等\n\n\n方案优势\n====\n\n\n1. 用户源站在DDoS IP高防和Web应用防火墙之后，起到隐藏源站IP的作用。\n2. CNAME接入，配置简单，减少运维人员工作。\n\n\n",
3    "title": "DDoS IP高防结合Web应用防火墙方案说明",
4    "product": "DDoS IP高防",
5    "url": "https://docs.jdcloud.com/cn/anti-ddos-pro/anti-ddos-pro-and-waf"
6}

每条数据是一个包含四个字段的json，这四个字段分别是"content"：文档内容；"title"：文档标题；"product"：相关产品；"url"：文档在线地址

向量数据库的选择与Retriever实现

向量数据库是RAG系统的记忆中心。目前市面上开源的向量数据库很多，那个向量库比较好也是见仁见智。本项目中笔者选择则了clickhouse作为向量数据库。选择ck主要有一下几个方面的考虑：

• ck再langchain社区的集成实现比较好，入库比较平滑
• 向量查询支持sql，学习成本较低，上手容易
• 京东云有相关产品且有专业团队支持，用着放心

文档向量化及入库过程

为了简化文档向量化和检索过程，我们使用了longchain的Retriever工具集
首先将文档向量化，代码如下：

1from libs.jd_doc_json_loader import JD_DOC_Loader
2from langchain_community.document_loaders import DirectoryLoader
3
4root_dir = "/root/jd_docs"
5loader = DirectoryLoader(
6    '/root/jd_docs', glob="**/*.json", loader_cls=JD_DOC_Loader)
7docs = loader.load()

langchain 社区里并没有提供针对特定格式的装载器，为此，我们自定义了JD_DOC_Loader来实现加载过程

1import json
2import logging
3from pathlib import Path
4from typing import Iterator, Optional, Union
5
6from langchain_core.documents import Document
7
8from langchain_community.document_loaders.base import BaseLoader
9from langchain_community.document_loaders.helpers import detect_file_encodings
10
11logger = logging.getLogger(__name__)
12
13
14class JD_DOC_Loader(BaseLoader):
15    """Load text file.
16
17
18    Args:
19        file_path: Path to the file to load.
20
21        encoding: File encoding to use. If `None`, the file will be loaded
22        with the default system encoding.
23
24        autodetect_encoding: Whether to try to autodetect the file encoding
25            if the specified encoding fails.
26    """
27
28    def __init__(
29        self,
30        file_path: Union[str, Path],
31        encoding: Optional[str] = None,
32        autodetect_encoding: bool = False,
33    ):
34        """Initialize with file path."""
35        self.file_path = file_path
36        self.encoding = encoding
37        self.autodetect_encoding = autodetect_encoding
38
39    def lazy_load(self) -> Iterator[Document]:
40        """Load from file path."""
41        text = ""
42        from_url = ""
43        try:
44            with open(self.file_path, encoding=self.encoding) as f:
45                doc_data = json.load(f)
46                text = doc_data["content"]
47                title = doc_data["title"]
48                product = doc_data["product"]
49                from_url = doc_data["url"]
50
51                # text = f.read()
52        except UnicodeDecodeError as e:
53            if self.autodetect_encoding:
54                detected_encodings = detect_file_encodings(self.file_path)
55                for encoding in detected_encodings:
56                    logger.debug(f"Trying encoding: {encoding.encoding}")
57                    try:
58                        with open(self.file_path, encoding=encoding.encoding) as f:
59                            text = f.read()
60                        break
61                    except UnicodeDecodeError:
62                        continue
63            else:
64                raise RuntimeError(f"Error loading {self.file_path}") from e
65        except Exception as e:
66            raise RuntimeError(f"Error loading {self.file_path}") from e
67        # metadata = {"source": str(self.file_path)}
68        metadata = {"source": from_url, "title": title, "product": product}
69        yield Document(page_content=text, metadata=metadata)

以上代码功能主要是解析json文件，填充Document的page_content字段和metadata字段。

接下来使用langchain 的 clickhouse 向量工具集进行文档入库

1import langchain_community.vectorstores.clickhouse as clickhouse
2from langchain.embeddings import HuggingFaceEmbeddings
3
4model_kwargs = {"device": "cuda"}
5embeddings = HuggingFaceEmbeddings(
6    model_name="/root/models/moka-ai-m3e-large", model_kwargs=model_kwargs)
7
8settings = clickhouse.ClickhouseSettings(
9    table="jd_docs_m3e_with_url", username="default", password="xxxxxx", host="10.0.1.94")
10
11docsearch = clickhouse.Clickhouse.from_documents(
12    docs, embeddings, config=settings)

入库成功后，进行一下检验

1import langchain_community.vectorstores.clickhouse as clickhouse
2from langchain.embeddings import HuggingFaceEmbeddings
3
4model_kwargs = {"device": "cuda"}~~~~
5embeddings = HuggingFaceEmbeddings(
6    model_name="/root/models/moka-ai-m3e-large", model_kwargs=model_kwargs)
7
8settings = clickhouse.ClickhouseSettings(
9    table="jd_docs_m3e_with_url_splited", username="default", password="xxxx", host="10.0.1.94")
10ck_db = clickhouse.Clickhouse(embeddings, config=settings)
11ck_retriever = ck_db.as_retriever(
12    search_type="similarity_score_threshold", search_kwargs={'score_threshold': 0.9})
13ck_retriever.get_relevant_documents("如何创建mysql rds")

有了知识库以后，可以构建一个简单的restful 服务，我们这里使用fastapi做这个事儿

1from fastapi import FastAPI
2from pydantic import BaseModel
3from singleton_decorator import singleton
4from langchain_community.embeddings import HuggingFaceEmbeddings
5import langchain_community.vectorstores.clickhouse as clickhouse
6import uvicorn
7import json
8
9app = FastAPI()
10app = FastAPI(docs_url=None)
11app.host = "0.0.0.0"
12
13model_kwargs = {"device": "cuda"}
14embeddings = HuggingFaceEmbeddings(
15    model_name="/root/models/moka-ai-m3e-large", model_kwargs=model_kwargs)
16settings = clickhouse.ClickhouseSettings(
17    table="jd_docs_m3e_with_url_splited", username="default", password="xxxx", host="10.0.1.94")
18ck_db = clickhouse.Clickhouse(embeddings, config=settings)
19ck_retriever = ck_db.as_retriever(
20    search_type="similarity", search_kwargs={"k": 3})
21
22
23class question(BaseModel):
24    content: str
25
26
27@app.get("/")
28async def root():
29    return {"ok"}
30
31
32@app.post("/retriever")
33async def retriver(question: question):
34    global ck_retriever
35    result = ck_retriever.invoke(question.content)
36    return result
37
38
39if __name__ == '__main__':
40    uvicorn.run(app='retriever_api:app', host="0.0.0.0",
41                port=8000, reload=True)

返回结构大概这样：

1[
2  {
3    "page_content": "云缓存 Redis--Redis迁移解决方案\n###RedisSyncer 操作步骤\n####数据校验\n```\nwget   https://github.com/TraceNature/rediscompare/releases/download/v1.0.0/rediscompare-1.0.0-linux-amd64.tar.gz\nrediscompare compare single2single --saddr \"10.0.1.101:6479\" --spassword \"redistest0102\" --taddr \"10.0.1.102:6479\" --tpassword  \"redistest0102\" --comparetimes 3\n\n```  \n**Github 地址：** [https://github.com/TraceNature/redissyncer-server](\"https://github.com/TraceNature/redissyncer-server\")",
4    "metadata": {
5      "product": "云缓存 Redis",
6      "source": "https://docs.jdcloud.com/cn/jcs-for-redis/doc-2",
7      "title": "Redis迁移解决方案"
8    },
9    "type": "Document"
10  },
11  {
12    "page_content": "云缓存 Redis--Redis迁移解决方案\n###RedisSyncer 操作步骤\n####数据校验\n```\nwget   https://github.com/TraceNature/rediscompare/releases/download/v1.0.0/rediscompare-1.0.0-linux-amd64.tar.gz\nrediscompare compare single2single --saddr \"10.0.1.101:6479\" --spassword \"redistest0102\" --taddr \"10.0.1.102:6479\" --tpassword  \"redistest0102\" --comparetimes 3\n\n```  \n**Github 地址：** [https://github.com/TraceNature/redissyncer-server](\"https://github.com/TraceNature/redissyncer-server\")",
13    "metadata": {
14      "product": "云缓存 Redis",
15      "source": "https://docs.jdcloud.com/cn/jcs-for-redis/doc-2",
16      "title": "Redis迁移解决方案"
17    },
18    "type": "Document"
19  },
20  {
21    "page_content": "云缓存 Redis--Redis迁移解决方案\n###RedisSyncer 操作步骤\n####数据校验\n```\nwget   https://github.com/TraceNature/rediscompare/releases/download/v1.0.0/rediscompare-1.0.0-linux-amd64.tar.gz\nrediscompare compare single2single --saddr \"10.0.1.101:6479\" --spassword \"redistest0102\" --taddr \"10.0.1.102:6479\" --tpassword  \"redistest0102\" --comparetimes 3\n\n```  \n**Github 地址：** [https://github.com/TraceNature/redissyncer-server](\"https://github.com/TraceNature/redissyncer-server\")",
22    "metadata": {
23      "product": "云缓存 Redis",
24      "source": "https://docs.jdcloud.com/cn/jcs-for-redis/doc-2",
25      "title": "Redis迁移解决方案"
26    },
27    "type": "Document"
28  }
29]

返回一个向量距离最小的list

结合模型和prompt，回答问题

为了节约算力资源，我们选择qwen 1.8B模型，一张v100卡刚好可以容纳一个qwen模型和一个m3e-large embedding 模型

• answer 服务

1from fastapi import FastAPI
2from pydantic import BaseModel
3from langchain_community.llms import VLLM
4from transformers import AutoTokenizer
5from langchain.prompts import PromptTemplate
6import requests
7import uvicorn
8import json
9import logging
10
11app = FastAPI()
12app = FastAPI(docs_url=None)
13app.host = "0.0.0.0"
14
15logger = logging.getLogger()
16logger.setLevel(logging.INFO)
17to_console = logging.StreamHandler()
18logger.addHandler(to_console)
19
20
21# load model
22# model_name = "/root/models/Llama3-Chinese-8B-Instruct"
23model_name = "/root/models/Qwen1.5-1.8B-Chat"
24tokenizer = AutoTokenizer.from_pretrained(model_name)
25llm_llama3 = VLLM(
26    model=model_name,
27    tokenizer=tokenizer,
28    task="text-generation",
29    temperature=0.2,
30    do_sample=True,
31    repetition_penalty=1.1,
32    return_full_text=False,
33    max_new_tokens=900,
34)
35
36# prompt
37prompt_template = """
38你是一个云技术专家
39使用以下检索到的Context回答问题。
40如果不知道答案，就说不知道。
41用中文回答问题。
42Question: {question}
43Context: {context}
44Answer: 
45"""
46
47prompt = PromptTemplate(
48    input_variables=["context", "question"],
49    template=prompt_template,
50)
51
52
53def get_context_list(q: str):
54    url = "http://10.0.0.7:8000/retriever"
55    payload = {"content": q}
56    res = requests.post(url, json=payload)
57    return res.text
58
59
60class question(BaseModel):
61    content: str
62
63
64@app.get("/")
65async def root():
66    return {"ok"}
67
68
69@app.post("/answer")
70async def answer(q: question):
71    logger.info("invoke!!!")
72    global prompt
73    global llm_llama3
74    context_list_str = get_context_list(q.content)
75
76    context_list = json.loads(context_list_str)
77    context = ""
78    source_list = []
79
80    for context_json in context_list:
81        context = context+context_json["page_content"]
82        source_list.append(context_json["metadata"]["source"])
83    p = prompt.format(context=context, question=q.content)
84    answer = llm_llama3(p)
85    result = {
86        "answer": answer,
87        "sources": source_list
88    }
89    return result
90
91
92if __name__ == '__main__':
93    uvicorn.run(app='retriever_api:app', host="0.0.0.0",
94                port=8888, reload=True)

代码通过使用Retriever接口查找与问题相似的文档，作为context组合prompt推送给模型生成答案。
主要服务就绪后可以开始画一张脸了，使用gradio做个简易对话界面

• gradio 服务

1import json
2import gradio as gr
3import requests
4
5
6def greet(name, intensity):
7    return "Hello, " + name + "!" * int(intensity)
8
9
10def answer(question):
11    url = "http://127.0.0.1:8888/answer"
12    payload = {"content": question}
13    res = requests.post(url, json=payload)
14    res_json = json.loads(res.text)
15    return [res_json["answer"], res_json["sources"]]
16
17
18demo = gr.Interface(
19    fn=answer,
20    # inputs=["text", "slider"],
21    inputs=[gr.Textbox(label="question", lines=5)],
22    # outputs=[gr.TextArea(label="answer", lines=5),
23    #          gr.JSON(label="urls", value=list)]
24    outputs=[gr.Markdown(label="answer"),
25             gr.JSON(label="urls", value=list)]
26)
27
28
29demo.launch(server_name="0.0.0.0")