大语言模型-中文Langchain
中文Langchain
使用chatGLM6b + langchain实现本地化知识库检索与智能答案生成
https://github.com/yanqiangmiffy/Chinese-LangChain
配置
class LangChainCFG:
llm_model_name = 'chatglm-6b' # 本地模型文件 or huggingface远程仓库
embedding_model_name = 'text2vec-large-chinese' # 检索模型文件 or huggingface远程仓库
vector_store_path = '.'
docs_path = './docs'
向量近邻检索
文本进行向量化后存入数据库，还不清楚langchain包里的FAISS做了哪些改变
class SourceService(object):
def __init__(self, config):
self.vector_store = None
self.config = config
self.embeddings = HuggingFaceEmbeddings(model_name=self.config.embedding_model_name)
self.docs_path = self.config.docs_path
self.vector_store_path = self.config.vector_store_path
def init_source_vector(self):
"""
初始化本地知识库向量
:return:
"""
docs = []
for doc in os.listdir(self.docs_path):
if doc.endswith('.txt'):
print(doc)
loader = UnstructuredFileLoader(f'{self.docs_path}/{doc}', mode="elements")
doc = loader.load()
docs.extend(doc)
self.vector_store = FAISS.from_documents(docs, self.embeddings)
self.vector_store.save_local(self.vector_store_path)
def add_document(self, document_path):
loader = UnstructuredFileLoader(document_path, mode="elements")
doc = loader.load()
self.vector_store.add_documents(doc)
self.vector_store.save_local(self.vector_store_path)
def load_vector_store(self, path):
if path is None:
self.vector_store = FAISS.load_local(self.vector_store_path, self.embeddings)
else:
self.vector_store = FAISS.load_local(path, self.embeddings)
return self.vector_store
def search_web(self, query):
SESSION.proxies = {
"http": f"socks5h://localhost:7890",
"https": f"socks5h://localhost:7890"
}
results = ddg(query)
web_content = ''
if results:
for result in results:
web_content += result['body']
return web_content
chatGLM
调用chatGLM
from typing import List, Optional
from langchain.llms.base import LLM
from langchain.llms.utils import enforce_stop_tokens
from transformers import AutoModel, AutoTokenizer
class ChatGLMService(LLM):
max_token: int = 10000
temperature: float = 0.1
top_p = 0.9
history = []
tokenizer: object = None
model: object = None
def __init__(self):
super().__init__()
@property
def _llm_type(self) -> str:
return "ChatGLM"
def _call(self,
prompt: str,
stop: Optional[List[str]] = None) -> str:
response, _ = self.model.chat(
self.tokenizer,
prompt,
history=self.history,
max_length=self.max_token,
temperature=self.temperature,
)
if stop is not None:
response = enforce_stop_tokens(response, stop)
self.history = self.history + [[None, response]]
return response
def load_model(self,
model_name_or_path: str = "THUDM/chatglm-6b"):
self.tokenizer = AutoTokenizer.from_pretrained(
model_name_or_path,
trust_remote_code=True
)
self.model = AutoModel.from_pretrained(model_name_or_path, trust_remote_code=True).half().cuda()
self.model=self.model.eval()
langchain
道理还是一样，把搜索的结果包装成prompt之后调用LLM
class LangChainApplication(object):
def __init__(self, config):
self.config = config
self.llm_service = ChatGLMService()
self.llm_service.load_model(model_name_or_path=self.config.llm_model_name)
self.source_service = SourceService(config)
if self.config.kg_vector_stores is None:
print("init a source vector store")
self.source_service.init_source_vector()
else:
print("load zh_wikipedia source vector store ")
try:
self.source_service.load_vector_store(self.config.kg_vector_stores['初始化知识库'])
except Exception as e:
self.source_service.init_source_vector()
def get_knowledge_based_answer(self, query,
history_len=5,
temperature=0.1,
top_p=0.9,
top_k=4,
web_content='',
chat_history=[]):
if web_content:
prompt_template = f"""基于以下已知信息，简洁和专业的来回答用户的问题。
如果无法从中得到答案，请说 "根据已知信息无法回答该问题" 或 "没有提供足够的相关信息"，不允许在答案中添加编造成分，答案请使用中文。
已知网络检索内容：{web_content}""" + """
已知内容:
{context}
问题:
{question}"""
else:
prompt_template = """基于以下已知信息，简洁和专业的来回答用户的问题。
如果无法从中得到答案，请说 "根据已知信息无法回答该问题" 或 "没有提供足够的相关信息"，不允许在答案中添加编造成分，答案请使用中文。
已知内容:
{context}
问题:
{question}"""
prompt = PromptTemplate(template=prompt_template,
input_variables=["context", "question"])
self.llm_service.history = chat_history[-history_len:] if history_len > 0 else []
self.llm_service.temperature = temperature
self.llm_service.top_p = top_p
knowledge_chain = RetrievalQA.from_llm(
llm=self.llm_service,
retriever=self.source_service.vector_store.as_retriever(
search_kwargs={"k": top_k}),
prompt=prompt)
knowledge_chain.combine_documents_chain.document_prompt = PromptTemplate(
input_variables=["page_content"], template="{page_content}")
knowledge_chain.return_source_documents = True
result = knowledge_chain({"query": query})
return result
# if __name__ == '__main__':
# config = LangChainCFG()
# application = LangChainApplication(config)
# result = application.get_knowledge_based_answer('马保国是谁')
# print(result)
# application.source_service.add_document('/home/searchgpt/yq/Knowledge-ChatGLM/docs/added/马保国.txt')
# result = application.get_knowledge_based_answer('马保国是谁')
# print(result)
web search
from duckduckgo_search import ddg
from duckduckgo_search.utils import SESSION
SESSION.proxies = {
"http": f"socks5h://localhost:7890",
"https": f"socks5h://localhost:7890"
}
r = ddg("马保国")
print(r[:2])
"""
[{'title': '马保国 - 维基百科，自由的百科全书', 'href': 'https://zh.wikipedia.org/wiki/%E9%A9%AC%E4%BF%9D%E5%9B%BD', 'body': '马保国（1951年 — ） ，男，籍贯 山东 临沂，出生及长大于河南，中国大陆太极拳师，自称"浑元形意太极门掌门人" 。 马保国因2017年约战mma格斗家徐晓冬首次出现
大众视野中。 2020年5月，马保国在对阵民间武术爱好者王庆民的比赛中，30秒内被连续高速击倒三次，此事件成为了持续多日的社交 ...'}, {'title': '馬保國的主页 - 抖音', 'href': 'https://www.douyin.com/user/MS4wLjABAAAAW0E1ziOvxgUh3VVv5FE6xmoo3w5WtZalfphYZKj4mCg', 'body': '6.3万. #马马国教扛打功 最近有几个人模芳我动作，很危险啊，不可以的，朋友们不要受伤了。. 5.3万. #马保国直播带货榜第一 朋友们周末愉快，本周六早上湿点，我本人在此号进行第一次带货直播，活到老，学到老，越活越年轻。. 7.0万. #马保国击破红牛罐 昨天 ...'}]
"""
webui
采用transformers的gradio
huggingface space
- https://huggingface.co/spaces/launch
- https://huggingface.co/docs/hub/spaces
- 可以使用github actions 同步：https://github.com/marketplace/actions/sync-with-hugging-face-hub