Spaces:
Sleeping
Sleeping
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,125 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
from langchain_community.document_loaders import UnstructuredMarkdownLoader
|
3 |
+
from langchain.text_splitter import RecursiveCharacterTextSplitter
|
4 |
+
from langchain_core.documents import Document
|
5 |
+
from langchain_huggingface import HuggingFaceEmbeddings
|
6 |
+
from langchain.vectorstores import FAISS
|
7 |
+
from langchain_community.llms import HuggingFaceHub
|
8 |
+
from langchain.prompts import ChatPromptTemplate
|
9 |
+
from dotenv import load_dotenv
|
10 |
+
import os
|
11 |
+
|
12 |
+
# Загрузка переменных окружения
|
13 |
+
load_dotenv()
|
14 |
+
|
15 |
+
DATA_PATH = ""
|
16 |
+
PROMPT_TEMPLATE = """
|
17 |
+
Ответь на вопрос, используя только следующий контекст:
|
18 |
+
{context}
|
19 |
+
---
|
20 |
+
Ответь на вопрос на основе приведенного контекста: {question}
|
21 |
+
"""
|
22 |
+
|
23 |
+
# Глобальная переменная для статуса
|
24 |
+
status_message = "Инициализация..."
|
25 |
+
|
26 |
+
def initialize_vectorstore():
|
27 |
+
global status_message
|
28 |
+
try:
|
29 |
+
status_message = "Загрузка и обработка документов..."
|
30 |
+
documents = load_documents()
|
31 |
+
chunks = split_text(documents)
|
32 |
+
|
33 |
+
status_message = "Создание векторной базы..."
|
34 |
+
vectorstore = save_to_faiss(chunks)
|
35 |
+
|
36 |
+
status_message = "База данных готова к использованию."
|
37 |
+
return vectorstore
|
38 |
+
|
39 |
+
except Exception as e:
|
40 |
+
status_message = f"Ошибка инициализации: {str(e)}"
|
41 |
+
raise
|
42 |
+
|
43 |
+
def generate_data_store():
|
44 |
+
documents = load_documents()
|
45 |
+
if documents:
|
46 |
+
chunks = split_text(documents)
|
47 |
+
return save_to_faiss(chunks)
|
48 |
+
|
49 |
+
def load_documents():
|
50 |
+
file_path = os.path.join(DATA_PATH, "pl250320252.md")
|
51 |
+
if not os.path.exists(file_path):
|
52 |
+
raise FileNotFoundError(f"Файл {file_path} не найден")
|
53 |
+
loader = UnstructuredMarkdownLoader(file_path)
|
54 |
+
return loader.load()
|
55 |
+
|
56 |
+
def split_text(documents: list[Document]):
|
57 |
+
text_splitter = RecursiveCharacterTextSplitter(
|
58 |
+
chunk_size=900,
|
59 |
+
chunk_overlap=300,
|
60 |
+
length_function=len,
|
61 |
+
add_start_index=True,
|
62 |
+
)
|
63 |
+
return text_splitter.split_documents(documents)
|
64 |
+
|
65 |
+
def save_to_faiss(chunks: list[Document]):
|
66 |
+
embeddings = HuggingFaceEmbeddings(
|
67 |
+
model_name="sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2",
|
68 |
+
model_kwargs={'device': 'cpu'},
|
69 |
+
encode_kwargs={'normalize_embeddings': True}
|
70 |
+
)
|
71 |
+
return FAISS.from_documents(chunks, embeddings)
|
72 |
+
|
73 |
+
def process_query(query_text: str, vectorstore):
|
74 |
+
if vectorstore is None:
|
75 |
+
return "База данных не инициализирована", []
|
76 |
+
|
77 |
+
try:
|
78 |
+
results = vectorstore.similarity_search_with_relevance_scores(query_text, k=3)
|
79 |
+
global status_message
|
80 |
+
status_message += f"\nНайдено {len(results)} результатов"
|
81 |
+
|
82 |
+
if not results:
|
83 |
+
return "Не найдено результатов.", []
|
84 |
+
|
85 |
+
context_text = "\n\n---\n\n".join([
|
86 |
+
f"Релевантность: {score:.2f}\n{doc.page_content}"
|
87 |
+
for doc, score in results
|
88 |
+
])
|
89 |
+
|
90 |
+
prompt_template = ChatPromptTemplate.from_template(PROMPT_TEMPLATE)
|
91 |
+
prompt = prompt_template.format(context=context_text, question=query_text)
|
92 |
+
|
93 |
+
model = HuggingFaceHub(
|
94 |
+
repo_id="google/flan-t5-small",
|
95 |
+
model_kwargs={"temperature": 0.5, "max_length": 512}
|
96 |
+
)
|
97 |
+
response_text = model.predict(prompt)
|
98 |
+
|
99 |
+
sources = list(set([doc.metadata.get("source", "") for doc, _ in results]))
|
100 |
+
return response_text, sources
|
101 |
+
|
102 |
+
except Exception as e:
|
103 |
+
return f"Ошибка обработки запроса: {str(e)}", []
|
104 |
+
|
105 |
+
def chat_interface(query_text):
|
106 |
+
global status_message
|
107 |
+
try:
|
108 |
+
vectorstore = initialize_vectorstore()
|
109 |
+
response, sources = process_query(query_text, vectorstore)
|
110 |
+
full_response = f"{status_message}\n\nОтвет: {response}\n\nИсточники: {', '.join(sources) if sources else 'Нет источников'}"
|
111 |
+
return full_response
|
112 |
+
except Exception as e:
|
113 |
+
return f"Критическая ошибка: {str(e)}"
|
114 |
+
|
115 |
+
# Интерфейс Gradio
|
116 |
+
interface = gr.Interface(
|
117 |
+
fn=chat_interface,
|
118 |
+
inputs=gr.Textbox(lines=2, placeholder="Введите ваш вопрос здесь..."),
|
119 |
+
outputs="text",
|
120 |
+
title="Чат с документами",
|
121 |
+
description="Задайте вопрос, и я отвечу на основе загруженных документов."
|
122 |
+
)
|
123 |
+
|
124 |
+
if __name__ == "__main__":
|
125 |
+
interface.launch()
|