Update app.py
Browse files
app.py
CHANGED
@@ -1,53 +1,13 @@
|
|
1 |
-
import
|
2 |
-
|
3 |
-
|
4 |
-
import
|
5 |
import os
|
6 |
-
import tensorflow_text
|
7 |
-
from sklearn.neighbors import NearestNeighbors
|
8 |
-
import gradio as gr
|
9 |
-
import requests
|
10 |
-
import json
|
11 |
-
import fitz
|
12 |
-
|
13 |
-
#这里填写调用openai需要的密钥
|
14 |
-
openai.api_key = '9481961416fa4c8e883047c5679cf971'
|
15 |
-
openai.api_base = 'https://demopro-oai-we2.openai.azure.com/'
|
16 |
-
openai.api_type = 'azure'
|
17 |
-
openai.api_version = '2022-12-01'
|
18 |
-
|
19 |
-
#将嵌套的列表展平
|
20 |
-
def flatten(_2d_list):
|
21 |
-
flat_list = []
|
22 |
-
for element in _2d_list:
|
23 |
-
if type(element) is list:
|
24 |
-
for item in element:
|
25 |
-
flat_list.append(item)
|
26 |
-
else:
|
27 |
-
flat_list.append(element)
|
28 |
-
return flat_list
|
29 |
-
|
30 |
|
31 |
def preprocess(text):
|
32 |
-
text = text.replace('\n', '
|
33 |
-
text = re.sub('\s+', ' ', text)
|
34 |
return text
|
35 |
|
36 |
-
#将pdf文档按段落分
|
37 |
-
# def pdf_to_text(path):
|
38 |
-
# doc = pdfplumber.open(path)
|
39 |
-
# pages = doc.pages
|
40 |
-
# text_list=[]
|
41 |
-
# for page,d in enumerate(pages):
|
42 |
-
# d=d.extract_text()
|
43 |
-
# d=preprocess(d)
|
44 |
-
# text_list.append(d)
|
45 |
-
# doc.close()
|
46 |
-
|
47 |
-
# return text_list
|
48 |
-
|
49 |
-
|
50 |
-
|
51 |
def pdf_to_text(path, start_page=1, end_page=None):
|
52 |
doc = fitz.open(path)
|
53 |
total_pages = doc.page_count
|
@@ -59,125 +19,53 @@ def pdf_to_text(path, start_page=1, end_page=None):
|
|
59 |
|
60 |
for i in range(start_page - 1, end_page):
|
61 |
text = doc.load_page(i).get_text("text")
|
62 |
-
text = preprocess(text)
|
63 |
text_list.append(text)
|
64 |
|
65 |
doc.close()
|
66 |
return text_list
|
67 |
|
|
|
|
|
|
|
|
|
68 |
|
69 |
-
|
70 |
-
|
71 |
-
|
72 |
-
chunks = []
|
73 |
-
|
74 |
-
for idx, words in enumerate(text_toks):
|
75 |
-
for i in range(0, len(words), word_length):
|
76 |
-
chunk = words[i : i + word_length]
|
77 |
-
if (
|
78 |
-
(i + word_length) > len(words)
|
79 |
-
and (len(chunk) < word_length)
|
80 |
-
and (len(text_toks) != (idx + 1))
|
81 |
-
):
|
82 |
-
text_toks[idx + 1] = chunk + text_toks[idx + 1]
|
83 |
-
continue
|
84 |
-
chunk = ' '.join(chunk).strip()
|
85 |
-
chunk = f'[Page no. {idx+start_page}]' + ' ' + '"' + chunk + '"'
|
86 |
-
chunks.append(chunk)
|
87 |
-
return chunks
|
88 |
|
89 |
-
|
90 |
-
history=text_to_chunks(history,start_page=1)
|
91 |
|
92 |
|
93 |
-
def
|
94 |
-
|
95 |
-
|
96 |
-
|
97 |
-
|
98 |
-
|
99 |
-
|
100 |
-
|
101 |
-
def __init__(self):
|
102 |
-
#类初始化,使用google公司的多语言语句编码,第一次运行时需要十几分钟的时间下载
|
103 |
-
self.use =hub.load('https://tfhub.dev/google/universal-sentence-encoder-multilingual/3')
|
104 |
-
self.fitted = False
|
105 |
-
|
106 |
-
def get_text_embedding(self, texts, batch=1000):
|
107 |
-
embeddings = []
|
108 |
-
for i in range(0, len(texts), batch):
|
109 |
-
text_batch = texts[i : (i + batch)]
|
110 |
-
emb_batch = self.use(text_batch)
|
111 |
-
embeddings.append(emb_batch)
|
112 |
-
embeddings = np.vstack(embeddings)
|
113 |
-
return embeddings
|
114 |
-
|
115 |
-
|
116 |
-
|
117 |
-
#K近邻算法,找到与问题最相似的 k 个段落,这里的 k 即n_neighbors=10
|
118 |
-
def fit(self, data, batch=1000, n_neighbors=5):
|
119 |
-
self.data = data
|
120 |
-
self.embeddings = self.get_text_embedding(data, batch=batch)
|
121 |
-
n_neighbors = min(n_neighbors, len(self.embeddings))
|
122 |
-
self.nn = NearestNeighbors(n_neighbors=n_neighbors)
|
123 |
-
self.nn.fit(self.embeddings)
|
124 |
-
self.fitted = True
|
125 |
-
|
126 |
-
#定义了该方法后,实例就可以被当作函数调用,text参数即用户提出的问题,inp_emb为其转化成的向量
|
127 |
-
def __call__(self, text, return_data=True):
|
128 |
-
inp_emb = self.use([text])
|
129 |
|
130 |
|
131 |
-
|
132 |
-
|
133 |
-
|
134 |
-
|
135 |
-
|
136 |
-
|
137 |
-
|
138 |
-
|
139 |
-
#openai的api接口,engine参数为我们选择的大语言模型,prompt即提示词
|
140 |
-
def generate_text(prompt, engine="text-davinci-003"):
|
141 |
-
completions = openai.Completion.create(
|
142 |
-
engine=engine,
|
143 |
-
prompt=prompt,
|
144 |
-
max_tokens=512,
|
145 |
-
n=1,
|
146 |
-
stop=None,
|
147 |
-
temperature=0.7,
|
148 |
-
)
|
149 |
-
message = completions.choices[0].text
|
150 |
-
return message
|
151 |
|
152 |
|
153 |
def generate_answer(question):
|
154 |
-
|
155 |
-
|
156 |
-
|
157 |
-
|
158 |
-
|
159 |
-
#把匹配到的段落加进提示词
|
160 |
-
for c in topn_chunks:
|
161 |
-
prompt += c + '\n\n'
|
162 |
-
|
163 |
-
#提示词
|
164 |
-
prompt += '''
|
165 |
-
Instructions: 如果搜索结果中找不到相关信息,只需要回答'未在该文档中找到相关信息'。
|
166 |
-
如果找到了相关信息,请使用中文回答,回答尽量精确简洁。并在句子的末尾使用[页码]符号引用每个参考文献(每个结果的开头都有这个编号)
|
167 |
-
如果不确定答案是否正确,就仅给出相似段落的来源,不要回复错误的答案。
|
168 |
-
\n\nQuery: {question}\nAnswer:
|
169 |
'''
|
170 |
-
|
171 |
-
|
172 |
-
answer =
|
173 |
return answer
|
174 |
-
|
175 |
-
|
176 |
-
recommender = SemanticSearch()
|
177 |
-
recommender.fit(history)
|
178 |
-
|
179 |
-
|
180 |
-
#以下为web客户端搭建,运行后产生客户端界面
|
181 |
def ask_api(question):
|
182 |
|
183 |
if question.strip() == '':
|
@@ -185,29 +73,24 @@ def ask_api(question):
|
|
185 |
|
186 |
return generate_answer(question)
|
187 |
|
188 |
-
title = 'Chat With Statistical Learning'
|
189 |
-
description = """ 该机器人将以Trevor Hastie等人所著的The Elements of Statistical Learning Data Mining, Inference, and Prediction
|
190 |
-
(即我们上课所用的课本)为主题回答你的问题,如果所问问题与书的内容无关,将会返回"未在该文档中找到相关信息"
|
191 |
-
"""
|
192 |
|
193 |
-
with gr.Blocks() as demo:
|
194 |
-
gr.Markdown(f'<center><h1>{title}</h1></center>')
|
195 |
-
gr.Markdown(description)
|
196 |
|
197 |
-
|
198 |
-
|
199 |
-
|
200 |
-
|
201 |
-
|
|
|
|
|
|
|
202 |
|
203 |
-
|
204 |
-
|
|
|
|
|
|
|
|
|
|
|
205 |
|
206 |
-
btn.click(
|
207 |
-
ask_api,
|
208 |
-
inputs=[question],
|
209 |
-
outputs=[answer]
|
210 |
-
)
|
211 |
|
212 |
-
#参数share=True会产生一个公开网页,别人可以通过访问该网页使用你的模型,前提是你需要正在运行这段代码(将自己的电脑当作服务器)
|
213 |
demo.launch()
|
|
|
1 |
+
from langchain.embeddings import OpenAIEmbeddings
|
2 |
+
from langchain.vectorstores import Chroma
|
3 |
+
from langchain.llms import OpenAI
|
4 |
+
from langchain.chains.question_answering import load_qa_chain
|
5 |
import os
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
6 |
|
7 |
def preprocess(text):
|
8 |
+
text = text.replace('\n', '')
|
|
|
9 |
return text
|
10 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
11 |
def pdf_to_text(path, start_page=1, end_page=None):
|
12 |
doc = fitz.open(path)
|
13 |
total_pages = doc.page_count
|
|
|
19 |
|
20 |
for i in range(start_page - 1, end_page):
|
21 |
text = doc.load_page(i).get_text("text")
|
|
|
22 |
text_list.append(text)
|
23 |
|
24 |
doc.close()
|
25 |
return text_list
|
26 |
|
27 |
+
def law_split(path,name):
|
28 |
+
text_list=pdf_to_text(path)
|
29 |
+
text= ''.join(text_list)
|
30 |
+
text_split=re.split(r'第.+条\s',text)[1:]
|
31 |
|
32 |
+
for index, text in enumerate(text_split):
|
33 |
+
text=preprocess(text)
|
34 |
+
text_split[index]=f'《中华人民共和国{name}》 第{index+1}条 '+text
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
35 |
|
36 |
+
return text_split
|
|
|
37 |
|
38 |
|
39 |
+
def folder_read(path):
|
40 |
+
text_list=[]
|
41 |
+
paths=os.listdir(path)
|
42 |
+
for file in paths:
|
43 |
+
name=file.split('.')[0]
|
44 |
+
suffix=file.split('.')[-1]
|
45 |
+
if suffix=='pdf':
|
46 |
+
text_list+=law_split(f'{path}/{file}',name)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
47 |
|
48 |
|
49 |
+
return text_list
|
50 |
+
|
51 |
+
|
52 |
+
text_list=folder_read('laws')
|
53 |
+
embeddings = OpenAIEmbeddings()
|
54 |
+
vectordb = Chroma.from_texts(texts=text_list, embedding=embeddings)
|
55 |
+
llm = OpenAI(temperature=0.5,max_tokens=1024)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
56 |
|
57 |
|
58 |
def generate_answer(question):
|
59 |
+
prompt='''
|
60 |
+
请根据给出的法律条文回答问题,给出适当的法律建议。回答时要说出你引用的法律条文是第几条,并说出引用的每一条是哪部法律中的。
|
61 |
+
引用的法律条文不要超过两条,回答尽量简明扼要
|
62 |
+
如果问题与搜索结果无关,就仅回答"该问题与青少年法律无关"即可。
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
63 |
'''
|
64 |
+
most_relevant_texts = vectordb.max_marginal_relevance_search(question, k=5)
|
65 |
+
chain = load_qa_chain(llm)
|
66 |
+
answer = chain.run(input_documents=most_relevant_texts, question=question+prompt)
|
67 |
return answer
|
68 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
69 |
def ask_api(question):
|
70 |
|
71 |
if question.strip() == '':
|
|
|
73 |
|
74 |
return generate_answer(question)
|
75 |
|
|
|
|
|
|
|
|
|
76 |
|
|
|
|
|
|
|
77 |
|
78 |
+
title = '青少年法律科普问答'
|
79 |
+
description = """ 本bot旨在根据中华人民共和国的法律回答有关青少年的问题,目前囊括的法律有\n
|
80 |
+
《未成年人保护法》\n
|
81 |
+
《义务教育法》\n
|
82 |
+
《预防未成年人犯罪法》\n
|
83 |
+
《妇女儿童权益保护法》
|
84 |
+
"""
|
85 |
+
|
86 |
|
87 |
+
demo = gr.Interface(
|
88 |
+
title=title,
|
89 |
+
description=description,
|
90 |
+
fn=ask_api,
|
91 |
+
inputs=gr.Textbox(label="请输入与青少年法律相关的问题",lines=2),
|
92 |
+
outputs=gr.outputs.Textbox(label="参考回答"),
|
93 |
+
examples=[["未成年遭受网络欺凌该怎么办?"],['年满多少岁的儿童应当接受义务教育?'],['若发现离家出走的未成年人,应如何处理?']])
|
94 |
|
|
|
|
|
|
|
|
|
|
|
95 |
|
|
|
96 |
demo.launch()
|