app.py

import streamlit as st
import faiss
from sentence_transformers import SentenceTransformer
import pickle
import re
from transformers import pipeline
import torch


st.set_page_config(page_title = "Vietnamese Legal Question Answering System", page_icon= "🐧", layout="centered", initial_sidebar_state="collapsed")

with open("./static/styles.css") as f:
    st.markdown(f"<style>{f.read()}</style>", unsafe_allow_html=True)
    
st.markdown(
    """
    <h1 style="text-align: center;">Vietnamese Legal Question Answering System</h1>
    """, 
    unsafe_allow_html=True
)

with open('articles.pkl', 'rb') as file:
    articles = pickle.load(file)

index_loaded = faiss.read_index("sentence_embeddings_index_no_citation.faiss")


device = 0 if torch.cuda.is_available() else -1
if 'model_embedding' not in st.session_state:
    st.session_state.model_embedding = SentenceTransformer('bkai-foundation-models/vietnamese-bi-encoder', device = f"cuda:{device}")



# Replace this with your own checkpoint
model_checkpoint = "model"
question_answerer = pipeline("question-answering", model=model_checkpoint, device = device)
def question_answering(question):
    print(question)
    query_sentence = [question]
    query_embedding = st.session_state.model_embedding.encode(query_sentence)
    k = 1
    D, I = index_loaded.search(query_embedding.astype('float32'), k)  # D is distances, I is indices
    answer = [question_answerer(question=query_sentence[0], context=articles[I[0][i]], max_answer_len = 512) for i in range(k)]
    best_answer = max(answer, key=lambda x: x['score'])
    print(best_answer)
    if best_answer['score'] > 0.5:
        return best_answer['answer']
    return f"Tôi không chắc lắm nhưng có lẽ câu trả lời là: \n{best_answer['answer']}"

# if "messages" not in st.session_state:
#     st.session_state.messages = []

# for message in st.session_state.messages:
#     with st.chat_message(message["role"]):
#         st.markdown(message["content"])


def clean_answer(s):
    # Sử dụng regex để loại bỏ tất cả các ký tự đặc biệt ở cuối chuỗi
    return re.sub(r'[^aAàÀảẢáÁạẠăĂằẰẳẲẵẴắẮặẶâÂầẦẩẨẫẪấẤậẬbBcCdDđĐeEèÈẻẺẽẼéÉẹẸêÊềỀểỂễỄếẾệỆfFgGhHiIìÌỉỈĩĨíÍịỊjJkKlLmMnNoOòÒỏỎõÕóÓọỌôÔồỒổỔỗỖốỐộỘơƠờỜởỞỡỠớỚợỢpPqQrRsStTuUùÙủỦũŨúÚụỤưƯừỪửỬữỮứỨựỰvVwWxXyYỳỲỷỶỹỸýÝỵỴzZ0-9]+$', '', s)
    
# if prompt := st.chat_input("What is up?"):
#         st.session_state.messages.append({"role": "user", "content": prompt})
#         with st.chat_message("user"):
#             st.markdown(prompt)
#         response = clean_answer(question_answering(prompt))
#         with st.chat_message("assistant"):
#             st.markdown(response)
        
#         st.session_state.messages.append({"role": "assistant", "content": response})

if 'messages' not in st.session_state:
    st.session_state.messages = []

for message in st.session_state.messages:
    if message['role'] == 'assistant':
        avatar_class = "assistant-avatar"
        message_class = "assistant-message"
        avatar = './app/static/AI.png'
    else:
        avatar_class = "user-avatar"
        message_class = "user-message"
        avatar = './app/static/human.jpg'
    st.markdown(f"""
    <div class="{message_class}">
        <img src="{avatar}" class="{avatar_class}" />
        <div class="stMarkdown">{message['content']}</div>
    </div>
    """, unsafe_allow_html=True)

if prompt := st.chat_input(placeholder='Xin chào, tôi có thể giúp được gì cho bạn?'):
    st.markdown(f"""
    <div class="user-message">
        <img src="./app/static/human.jpg" class="user-avatar" />
        <div class="stMarkdown">{prompt}</div>
    </div>
    """, unsafe_allow_html=True)
    st.session_state.messages.append({'role': 'user', 'content': prompt})
    
    respond = clean_answer(question_answering(prompt))

    st.markdown(f"""
    <div class="assistant-message">
        <img src="./app/static/AI.png" class="assistant-avatar" />
        <div class="stMarkdown">{respond}</div>
    </div>
    """, unsafe_allow_html=True)
    st.session_state.messages.append({'role': 'assistant', 'content': respond})