Spaces:
Runtime error
Runtime error
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,266 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
|
3 |
+
import gradio as gr
|
4 |
+
import nltk
|
5 |
+
import sentence_transformers
|
6 |
+
import torch
|
7 |
+
from itertools import islice
|
8 |
+
from duckduckgo_search import ddg
|
9 |
+
from duckduckgo_search import DDGS
|
10 |
+
from langchain.chains import RetrievalQA
|
11 |
+
from langchain.document_loaders import UnstructuredFileLoader
|
12 |
+
from langchain.embeddings import JinaEmbeddings
|
13 |
+
from langchain.embeddings.huggingface import HuggingFaceEmbeddings
|
14 |
+
from langchain.prompts import PromptTemplate
|
15 |
+
from langchain.prompts.prompt import PromptTemplate
|
16 |
+
from langchain.vectorstores import FAISS
|
17 |
+
|
18 |
+
|
19 |
+
|
20 |
+
|
21 |
+
from chatllm import ChatLLM
|
22 |
+
from chinese_text_splitter import ChineseTextSplitter
|
23 |
+
|
24 |
+
nltk.data.path.append('./nltk_data')
|
25 |
+
|
26 |
+
embedding_model_dict = {
|
27 |
+
"ernie-tiny": "nghuyong/ernie-3.0-nano-zh",
|
28 |
+
"ernie-base": "nghuyong/ernie-3.0-base-zh",
|
29 |
+
"text2vec-base": "GanymedeNil/text2vec-base-chinese",
|
30 |
+
#"ViT-B-32": 'ViT-B-32::laion2b-s34b-b79k'
|
31 |
+
}
|
32 |
+
|
33 |
+
llm_model_dict = {
|
34 |
+
"ChatGLM-6B-int8": "THUDM/chatglm-6b-int8",
|
35 |
+
"ChatGLM-6B-int4": "THUDM/chatglm-6b-int4",
|
36 |
+
"ChatGLM-6b-int4-qe": "THUDM/chatglm-6b-int4-qe"
|
37 |
+
}
|
38 |
+
|
39 |
+
DEVICE = "cuda" if torch.cuda.is_available(
|
40 |
+
) else "mps" if torch.backends.mps.is_available() else "cpu"
|
41 |
+
|
42 |
+
|
43 |
+
def search_web(query):
|
44 |
+
web_content = ''
|
45 |
+
with DDGS() as ddgs:
|
46 |
+
results = ddgs.text(query, region='wt-wt', safesearch='Off');
|
47 |
+
if results:
|
48 |
+
for result in islice(results, 3):
|
49 |
+
web_content += result['body']
|
50 |
+
return web_content
|
51 |
+
|
52 |
+
|
53 |
+
def load_file(filepath):
|
54 |
+
if filepath.lower().endswith(".pdf"):
|
55 |
+
loader = UnstructuredFileLoader(filepath)
|
56 |
+
textsplitter = ChineseTextSplitter(pdf=True)
|
57 |
+
docs = loader.load_and_split(textsplitter)
|
58 |
+
else:
|
59 |
+
loader = UnstructuredFileLoader(filepath, mode="elements")
|
60 |
+
textsplitter = ChineseTextSplitter(pdf=False)
|
61 |
+
docs = loader.load_and_split(text_splitter=textsplitter)
|
62 |
+
return docs
|
63 |
+
|
64 |
+
|
65 |
+
def init_knowledge_vector_store(embedding_model, filepath):
|
66 |
+
if embedding_model == "ViT-B-32":
|
67 |
+
jina_auth_token = os.getenv('jina_auth_token')
|
68 |
+
embeddings = JinaEmbeddings(
|
69 |
+
jina_auth_token=jina_auth_token,
|
70 |
+
model_name=embedding_model_dict[embedding_model])
|
71 |
+
else:
|
72 |
+
embeddings = HuggingFaceEmbeddings(
|
73 |
+
model_name=embedding_model_dict[embedding_model], )
|
74 |
+
embeddings.client = sentence_transformers.SentenceTransformer(
|
75 |
+
embeddings.model_name, device=DEVICE)
|
76 |
+
|
77 |
+
docs = load_file(filepath)
|
78 |
+
|
79 |
+
vector_store = FAISS.from_documents(docs, embeddings)
|
80 |
+
return vector_store
|
81 |
+
|
82 |
+
|
83 |
+
def get_knowledge_based_answer(query,
|
84 |
+
large_language_model,
|
85 |
+
vector_store,
|
86 |
+
VECTOR_SEARCH_TOP_K,
|
87 |
+
web_content,
|
88 |
+
history_len,
|
89 |
+
temperature,
|
90 |
+
top_p,
|
91 |
+
chat_history=[]):
|
92 |
+
if web_content:
|
93 |
+
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。
|
94 |
+
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。
|
95 |
+
已知网络检索内容:{web_content}""" + """
|
96 |
+
已知内容:
|
97 |
+
{context}
|
98 |
+
问题:
|
99 |
+
{question}"""
|
100 |
+
else:
|
101 |
+
prompt_template = """基于以下已知信息,请简洁并专业地回答用户的问题。
|
102 |
+
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。另外,答案请使用中文。
|
103 |
+
|
104 |
+
已知内容:
|
105 |
+
{context}
|
106 |
+
|
107 |
+
问题:
|
108 |
+
{question}"""
|
109 |
+
prompt = PromptTemplate(template=prompt_template,
|
110 |
+
input_variables=["context", "question"])
|
111 |
+
chatLLM = ChatLLM()
|
112 |
+
chatLLM.history = chat_history[-history_len:] if history_len > 0 else []
|
113 |
+
if large_language_model == "ChatGPT":
|
114 |
+
chatLLM.model = OpenAI()
|
115 |
+
else:
|
116 |
+
chatLLM.load_model(
|
117 |
+
model_name_or_path=llm_model_dict[large_language_model])
|
118 |
+
chatLLM.temperature = temperature
|
119 |
+
chatLLM.top_p = top_p
|
120 |
+
|
121 |
+
knowledge_chain = RetrievalQA.from_llm(
|
122 |
+
llm=chatLLM,
|
123 |
+
retriever=vector_store.as_retriever(
|
124 |
+
search_kwargs={"k": VECTOR_SEARCH_TOP_K}),
|
125 |
+
prompt=prompt)
|
126 |
+
knowledge_chain.combine_documents_chain.document_prompt = PromptTemplate(
|
127 |
+
input_variables=["page_content"], template="{page_content}")
|
128 |
+
|
129 |
+
knowledge_chain.return_source_documents = True
|
130 |
+
|
131 |
+
result = knowledge_chain({"query": query})
|
132 |
+
return result
|
133 |
+
|
134 |
+
|
135 |
+
def clear_session():
|
136 |
+
return '', None
|
137 |
+
|
138 |
+
|
139 |
+
def predict(input,
|
140 |
+
large_language_model,
|
141 |
+
embedding_model,
|
142 |
+
file_obj,
|
143 |
+
VECTOR_SEARCH_TOP_K,
|
144 |
+
temperature,
|
145 |
+
top_p,
|
146 |
+
use_web,
|
147 |
+
history=None):
|
148 |
+
if history == None:
|
149 |
+
history = []
|
150 |
+
print(file_obj.name)
|
151 |
+
vector_store = init_knowledge_vector_store(embedding_model, file_obj.name)
|
152 |
+
if use_web == 'True':
|
153 |
+
web_content = search_web(query=input)
|
154 |
+
else:
|
155 |
+
web_content = ''
|
156 |
+
resp = get_knowledge_based_answer(
|
157 |
+
query=input,
|
158 |
+
large_language_model=large_language_model,
|
159 |
+
vector_store=vector_store,
|
160 |
+
VECTOR_SEARCH_TOP_K=VECTOR_SEARCH_TOP_K,
|
161 |
+
web_content=web_content,
|
162 |
+
chat_history=history,
|
163 |
+
history_len=history_len,
|
164 |
+
temperature=temperature,
|
165 |
+
top_p=top_p,
|
166 |
+
)
|
167 |
+
print(resp)
|
168 |
+
history.append((input, resp['result']))
|
169 |
+
return '', history, history
|
170 |
+
|
171 |
+
|
172 |
+
if __name__ == "__main__":
|
173 |
+
block = gr.Blocks()
|
174 |
+
with block as demo:
|
175 |
+
gr.Markdown("""<h1><center>LangChain-ChatLLM-Webui</center></h1>
|
176 |
+
<center><font size=3>
|
177 |
+
本项目基于LangChain和大型语言模型系列模型, 提供基于本地知识的自动问答应用. <br>
|
178 |
+
</center></font>
|
179 |
+
""")
|
180 |
+
with gr.Row():
|
181 |
+
with gr.Column(scale=1):
|
182 |
+
model_choose = gr.Accordion("模型选择")
|
183 |
+
with model_choose:
|
184 |
+
large_language_model = gr.Dropdown(
|
185 |
+
list(llm_model_dict.keys()),
|
186 |
+
label="large language model",
|
187 |
+
value="ChatGLM-6B-int4")
|
188 |
+
|
189 |
+
embedding_model = gr.Dropdown(list(
|
190 |
+
embedding_model_dict.keys()),
|
191 |
+
label="Embedding model",
|
192 |
+
value="text2vec-base")
|
193 |
+
|
194 |
+
file = gr.File(label='请上传知识库文件, 目前支持txt、docx、md格式',
|
195 |
+
file_types=['.txt', '.md', '.docx'])
|
196 |
+
|
197 |
+
use_web = gr.Radio(["True", "False"],
|
198 |
+
label="Web Search",
|
199 |
+
value="False")
|
200 |
+
model_argument = gr.Accordion("模型参数配置")
|
201 |
+
|
202 |
+
with model_argument:
|
203 |
+
|
204 |
+
VECTOR_SEARCH_TOP_K = gr.Slider(
|
205 |
+
1,
|
206 |
+
10,
|
207 |
+
value=6,
|
208 |
+
step=1,
|
209 |
+
label="vector search top k",
|
210 |
+
interactive=True)
|
211 |
+
|
212 |
+
# HISTORY_LEN = gr.Slider(0,
|
213 |
+
# 3,
|
214 |
+
# value=0,
|
215 |
+
# step=1,
|
216 |
+
# label="history len",
|
217 |
+
# interactive=True)
|
218 |
+
|
219 |
+
temperature = gr.Slider(0,
|
220 |
+
1,
|
221 |
+
value=0.01,
|
222 |
+
step=0.01,
|
223 |
+
label="temperature",
|
224 |
+
interactive=True)
|
225 |
+
top_p = gr.Slider(0,
|
226 |
+
1,
|
227 |
+
value=0.9,
|
228 |
+
step=0.1,
|
229 |
+
label="top_p",
|
230 |
+
interactive=True)
|
231 |
+
|
232 |
+
with gr.Column(scale=4):
|
233 |
+
chatbot = gr.Chatbot(label='ChatLLM').style(height=600)
|
234 |
+
message = gr.Textbox(label='请输入问题')
|
235 |
+
state = gr.State()
|
236 |
+
|
237 |
+
with gr.Row():
|
238 |
+
clear_history = gr.Button("🧹 清除历史对话")
|
239 |
+
send = gr.Button("🚀 发送")
|
240 |
+
|
241 |
+
send.click(predict,
|
242 |
+
inputs=[
|
243 |
+
message, large_language_model,
|
244 |
+
embedding_model, file, VECTOR_SEARCH_TOP_K,
|
245 |
+
HISTORY_LEN, temperature, top_p, use_web,
|
246 |
+
state
|
247 |
+
],
|
248 |
+
outputs=[message, chatbot, state])
|
249 |
+
clear_history.click(fn=clear_session,
|
250 |
+
inputs=[],
|
251 |
+
outputs=[chatbot, state],
|
252 |
+
queue=False)
|
253 |
+
|
254 |
+
message.submit(predict,
|
255 |
+
inputs=[
|
256 |
+
message, large_language_model,
|
257 |
+
embedding_model, file,
|
258 |
+
VECTOR_SEARCH_TOP_K, HISTORY_LEN,
|
259 |
+
temperature, top_p, use_web, state
|
260 |
+
],
|
261 |
+
outputs=[message, chatbot, state])
|
262 |
+
gr.Markdown("""提醒:<br>
|
263 |
+
1. 使用时请先上传自己的知识文件,并且文件中不含某些特殊字符,否则将返回error. <br>
|
264 |
+
2. 有任何使用问题,请通过[问题交流区](https://huggingface.co/spaces/thomas-yanxin/LangChain-ChatLLM/discussions)或[Github Issue区](https://github.com/thomas-yanxin/LangChain-ChatGLM-Webui/issues)进行反馈. <br>
|
265 |
+
""")
|
266 |
+
demo.queue().launch(server_name='0.0.0.0', share=False)
|