app.py

import streamlit as st
import torch
from transformers import AutoTokenizer
from semviqa.ser.qatc_model import QATCForQuestionAnswering
from semviqa.tvc.model import ClaimModelForClassification
from semviqa.ser.ser_eval import extract_evidence_tfidf_qatc
from semviqa.tvc.tvc_eval import classify_claim

# Load models with caching
@st.cache_resource()
def load_model(model_name, model_class):
    tokenizer = AutoTokenizer.from_pretrained(model_name)
    model = model_class.from_pretrained(model_name)
    return tokenizer, model

# UI Configuration
st.set_page_config(page_title="SemViQA Demo", layout="wide")

st.markdown("""
    <style>
        .big-title { font-size: 36px; font-weight: bold; color: #4A90E2; text-align: center; }
        .sub-title { font-size: 20px; color: #666; text-align: center; }
        .stButton>button { background-color: #4CAF50; color: white; font-size: 16px; width: 100%; border-radius: 8px; padding: 10px; }
        .stTextArea textarea { font-size: 16px; }
        .result-box { background-color: #f9f9f9; padding: 20px; border-radius: 10px; box-shadow: 0px 4px 8px rgba(0, 0, 0, 0.1); }
    </style>
""", unsafe_allow_html=True)

st.markdown("<p class='big-title'>🔍 SemViQA: Vietnamese Fact-Checking System</p>", unsafe_allow_html=True)
st.markdown("<p class='sub-title'>Enter a claim and context to verify its accuracy</p>", unsafe_allow_html=True)

# Sidebar - Configuration Settings
with st.sidebar.expander("⚙️ Settings", expanded=False):
    tfidf_threshold = st.slider("🔧 TF-IDF Threshold", 0.0, 1.0, 0.5, 0.01)
    length_ratio_threshold = st.slider("📏 Length Ratio Threshold", 0.1, 1.0, 0.5, 0.01)
    qatc_model_name = st.selectbox("🤖 QATC Model", ["xuandin/semviqa-qatc-vimrc-viwikifc"])
    bc_model_name = st.selectbox("🏷️ Binary Classification Model", ["xuandin/semviqa-bc"])
    tc_model_name = st.selectbox("📊 Three-Class Model", ["xuandin/semviqa-tc"])

# Load selected models
tokenizer_qatc, model_qatc = load_model(qatc_model_name, QATCForQuestionAnswering)
tokenizer_bc, model_bc = load_model(bc_model_name, ClaimModelForClassification)
tokenizer_tc, model_tc = load_model(tc_model_name, ClaimModelForClassification)

# User Input Fields
claim = st.text_area("✍️ Enter Claim", "Vietnam is a country in Southeast Asia.")
context = st.text_area("📖 Enter Context", "Vietnam is a country located in Southeast Asia, covering an area of over 331,000 km² with a population of more than 98 million people.")

if st.button("🔎 Verify"):
    # Extract evidence
    evidence = extract_evidence_tfidf_qatc(
        claim, context, model_qatc, tokenizer_qatc, "cuda" if torch.cuda.is_available() else "cpu",
        confidence_threshold=tfidf_threshold, length_ratio_threshold=length_ratio_threshold
    )
    
    # Claim Classification
    verdict = "NEI"
    prob3class, pred_tc = classify_claim(claim, evidence, model_tc, tokenizer_tc, "cuda" if torch.cuda.is_available() else "cpu")
    
    if pred_tc != 0:
        prob2class, pred_bc = classify_claim(claim, evidence, model_bc, tokenizer_bc, "cuda" if torch.cuda.is_available() else "cpu")
        verdict = "SUPPORTED" if pred_bc == 0 else "REFUTED" if prob2class > prob3class else ["NEI", "SUPPORTED", "REFUTED"][pred_tc]
    
    # Display Results
    st.markdown(f"""
        <div class='result-box'>
            <h3>📌 Result</h3>
            <p><strong>🔍 Evidence:</strong> {evidence}</p>
            <p><strong>✅ Verdict:</strong> {verdict}</p>
        </div>
    """, unsafe_allow_html=True)