app.py

import gradio as gr
from googleapiclient.discovery import build
import google.generativeai as genai
import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import re
import os
import numpy as np

GOOGLE_API_KEY = "AIzaSyDu0819TPX_Z1AcAT5xT1SNjjmb64PSc1I"
SEARCH_ENGINE_ID = "f34f8a4816771488b"
GEMINI_API_KEY = "AIzaSyAHPzJ_VjTX3gZLBV28d3sq97SdER2qfkc"
MODEL_PATH = "./vietnamese_fake_news_model"

genai.configure(api_key=GEMINI_API_KEY)

print("Loading the DistilBERT model we trained...")
try:
    if os.path.exists(MODEL_PATH):
        tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH)
        model = AutoModelForSequenceClassification.from_pretrained(MODEL_PATH)
        print("DistilBERT model loaded successfully!")
    else:
        print(f"Model directory '{MODEL_PATH}' not found!")
        print("Our custom model isn't available, trying a backup model...")
        try:
            tokenizer = AutoTokenizer.from_pretrained("distilbert-base-multilingual-cased")
            model = AutoModelForSequenceClassification.from_pretrained("distilbert-base-multilingual-cased", num_labels=2)
            print("Fallback DistilBERT model loaded successfully!")
        except Exception as fallback_error:
            print(f"Fallback model also failed: {fallback_error}")
            tokenizer = None
            model = None
except Exception as e:
    print(f"Error loading DistilBERT model: {e}")
    print("Something went wrong, trying the backup model...")
    try:
        tokenizer = AutoTokenizer.from_pretrained("distilbert-base-multilingual-cased")
        model = AutoModelForSequenceClassification.from_pretrained("distilbert-base-multilingual-cased", num_labels=2)
        print("Fallback DistilBERT model loaded successfully!")
    except Exception as fallback_error:
        print(f"Fallback model also failed: {fallback_error}")
        tokenizer = None
        model = None

CREDIBLE_SOURCES = {
    'vnexpress.net': 0.95,
    'tuoitre.vn': 0.95,
    'thanhnien.vn': 0.90,
    'dantri.com.vn': 0.90,
    'vietnamnet.vn': 0.85,
    'zing.vn': 0.85,
    'kenh14.vn': 0.80,
    'soha.vn': 0.80,
    'baotintuc.vn': 0.85,
    'nhandan.vn': 0.90,
    'laodong.vn': 0.85,
    'congan.com.vn': 0.90,
    'quochoi.vn': 0.95,
    'chinhphu.vn': 0.95,
    'moh.gov.vn': 0.90,
    'mofa.gov.vn': 0.90,
    'mard.gov.vn': 0.85,
    'moc.gov.vn': 0.85,
    'mof.gov.vn': 0.85,
    'mst.gov.vn': 0.85,
    'wikipedia.org': 0.95,
    'bbc.com': 0.95,
    'bbc.co.uk': 0.95,
    'cnn.com': 0.90,
    'reuters.com': 0.95,
    'ap.org': 0.95,
    'espn.com': 0.85,
    'fifa.com': 0.95,
    'nytimes.com': 0.90,
    'washingtonpost.com': 0.90,
    'theguardian.com': 0.90
}

def clean_text(text):
    """Clean up the text before feeding it to our model"""
    if not isinstance(text, str):
        text = str(text)
    text = re.sub(r'\s+', ' ', text.strip())
    if len(text) < 10:
        text = "Tin tức ngắn: " + text
    return text

def predict_with_distilbert(text):
    """Run the text through our trained DistilBERT model to get a prediction"""
    if model is None or tokenizer is None:
        return None, None, None, None
    
    try:
        clean_text_input = clean_text(text)
        inputs = tokenizer(
            clean_text_input,
            return_tensors="pt",
            truncation=True,
            padding=True,
            max_length=512
        )
        
        
        with torch.no_grad():
            outputs = model(**inputs)
            predictions = torch.nn.functional.softmax(outputs.logits, dim=-1)
        
        
        real_score = predictions[0][0].item()
        fake_score = predictions[0][1].item()
        
        
        if real_score > fake_score:
            prediction = "REAL"
            confidence = real_score
        else:
            prediction = "FAKE"
            confidence = fake_score
        
        return prediction, confidence, real_score, fake_score
        
    except Exception as e:
        print(f"DistilBERT prediction error: {e}")
        return None, None, None, None

def process_search_results(items):
    
    search_results = []
    for item in items:
        search_results.append({
            'title': item.get('title', ''),
            'snippet': item.get('snippet', ''),
            'link': item.get('link', '')
        })
    return search_results

def google_search_fallback(news_text):
    
    print("Using fallback search system...")
    
    mock_results = []
    
    
    if "Argentina" in news_text and "World Cup" in news_text:
        mock_results = [
            {
                'title': 'Argentina wins World Cup 2022 - FIFA Official',
                'snippet': 'Argentina defeated France in the 2022 World Cup final to win their third World Cup title.',
                'link': 'https://www.fifa.com/worldcup/news/argentina-wins-world-cup-2022'
            },
            {
                'title': 'World Cup 2022 Final: Argentina vs France - BBC Sport',
                'snippet': 'Argentina won the 2022 FIFA World Cup after defeating France in a thrilling final.',
                'link': 'https://www.bbc.com/sport/football/world-cup-2022'
            },
            {
                'title': 'Lionel Messi leads Argentina to World Cup victory - ESPN',
                'snippet': 'Lionel Messi finally won the World Cup as Argentina defeated France in Qatar 2022.',
                'link': 'https://www.espn.com/soccer/world-cup/story/argentina-messi-world-cup'
            }
        ]
    elif "COVID" in news_text or "covid" in news_text:
        mock_results = [
            {
                'title': 'COVID-19 Updates - World Health Organization',
                'snippet': 'Latest updates on COVID-19 pandemic from WHO official sources.',
                'link': 'https://www.who.int/emergencies/diseases/novel-coronavirus-2019'
            },
            {
                'title': 'COVID-19 Vietnam News - Ministry of Health',
                'snippet': 'Official COVID-19 updates from Vietnam Ministry of Health.',
                'link': 'https://moh.gov.vn/covid-19'
            }
        ]
    elif "Việt Nam" in news_text or "Vietnam" in news_text:
        mock_results = [
            {
                'title': 'Vietnam News - VnExpress',
                'snippet': 'Latest news from Vietnam covering politics, economy, and society.',
                'link': 'https://vnexpress.net'
            },
            {
                'title': 'Vietnam News - Tuổi Trẻ',
                'snippet': 'Vietnamese news and current events from Tuổi Trẻ newspaper.',
                'link': 'https://tuoitre.vn'
            }
        ]
    else:
        
        mock_results = [
            {
                'title': 'News Verification - Fact Check',
                'snippet': 'Fact-checking and news verification from reliable sources.',
                'link': 'https://www.factcheck.org'
            },
            {
                'title': 'News Analysis - Reuters',
                'snippet': 'Professional news analysis and reporting from Reuters.',
                'link': 'https://www.reuters.com'
            }
        ]
    
    print(f"Generated {len(mock_results)} mock search results")
    return mock_results



def google_search(news_text):
    """Search Google for information about the news, with backup options if it fails"""
    try:
        service = build("customsearch", "v1", developerKey=GOOGLE_API_KEY)
        
        
        search_queries = []
        
        
        if "Argentina" in news_text and "World Cup" in news_text:
            search_queries = [
                "Argentina World Cup 2022 champion winner",
                "Argentina vô địch World Cup 2022",
                "World Cup 2022 Argentina final"
            ]
        elif "COVID" in news_text or "covid" in news_text:
            search_queries = [
                "COVID-19 Vietnam news",
                "COVID Vietnam 2022 2023",
                "dịch COVID Việt Nam"
            ]
        else:
            
            vietnamese_words = re.findall(r'[À-ỹ]+', news_text)
            english_words = re.findall(r'[A-Za-z]+', news_text)
            numbers = re.findall(r'\d{4}', news_text)  # Years
            
            
            if english_words:
                search_queries.append(' '.join(english_words[:5]))
            if vietnamese_words:
                search_queries.append(' '.join(vietnamese_words[:5]))
            if numbers:
                search_queries.append(' '.join(english_words[:3] + numbers))
            
            
            keywords = re.findall(r'[A-Za-zÀ-ỹ]+|\b(?:19|20)\d{2}\b|\b\d{1,2}\b', news_text)
            search_queries.append(' '.join(keywords[:10]))
        
        
        for i, search_query in enumerate(search_queries):
            if not search_query.strip():
                continue
                
            print(f"Strategy {i+1}: Searching for '{search_query}'")
            
            result = service.cse().list(
                q=search_query,
                cx=SEARCH_ENGINE_ID,
                num=10
            ).execute()
            
            print(f"API Response keys: {list(result.keys())}")
            if 'searchInformation' in result:
                print(f"Total results: {result['searchInformation'].get('totalResults', 'Unknown')}")
            
            if 'items' in result and result['items']:
                print(f"Found {len(result['items'])} results with strategy {i+1}")
                return process_search_results(result['items'])
            else:
                print(f"No results with strategy {i+1}")
        
        
        print("All strategies failed, trying simple phrase search...")
        simple_query = news_text[:30]  # First 30 characters
        result = service.cse().list(
            q=simple_query,
            cx=SEARCH_ENGINE_ID,
            num=5
        ).execute()
        
        if 'items' in result and result['items']:
            print(f"Found {len(result['items'])} results with simple search")
            return process_search_results(result['items'])
        
        print("All search strategies failed, using fallback...")
        return google_search_fallback(news_text)
            
    except Exception as e:
        print(f"Google Search error: {e}")
        print(f"Error type: {type(e).__name__}")
        
        
        error_str = str(e).lower()
        if any(keyword in error_str for keyword in ["403", "blocked", "quota", "limit", "exceeded"]):
            print("Google Search API blocked/quota exceeded, using fallback...")
        elif "invalid" in error_str or "unauthorized" in error_str:
            print("API key issue, using fallback...")
        
        return google_search_fallback(news_text)

def analyze_sources(search_results):
    """Check how trustworthy the news sources are"""
    if not search_results:
        return 0.50, 0.20, "No sources found"
    
    
    credible_count = 0
    total_sources = len(search_results)
    
    for result in search_results:
        domain = result['link'].split('/')[2] if '//' in result['link'] else ''
        for source, credibility in CREDIBLE_SOURCES.items():
            if source in domain:
                credible_count += 1
                break
    
    source_credibility = credible_count / total_sources if total_sources > 0 else 0.50
    
    
    popularity_score = min(1.0, total_sources / 5.0)  # Normalize to 0-1
    
    # Create a summary of what we found
    if source_credibility > 0.7:
        credibility_text = f"High credibility: {credible_count}/{total_sources} sources from reputable outlets"
    elif source_credibility > 0.4:
        credibility_text = f"Medium credibility: {credible_count}/{total_sources} sources from reputable outlets"
    else:
        credibility_text = f"Low credibility: {credible_count}/{total_sources} sources from reputable outlets"
    
    return source_credibility, popularity_score, credibility_text

def analyze_source_support(news_text, search_results):
    """Check if the search results agree or disagree with the news"""
    if not search_results:
        return 0.5, "No sources to analyze"
    
    support_count = 0
    contradict_count = 0
    total_sources = len(search_results)
    
    # Look for years mentioned in the news
    import re
    news_years = re.findall(r'\b(20\d{2})\b', news_text)
    news_year = news_years[0] if news_years else None
    
    for result in search_results:
        title_snippet = (result.get('title', '') + ' ' + result.get('snippet', '')).lower()
        
        # See if the years match up
        if news_year:
            source_years = re.findall(r'\b(20\d{2})\b', title_snippet)
            if source_years and news_year not in source_years:
                contradict_count += 1
                continue
        
        # Look for words that suggest agreement or disagreement
        support_keywords = ['confirm', 'verify', 'true', 'accurate', 'correct', 'xác nhận', 'chính xác', 'đúng']
        contradict_keywords = ['false', 'fake', 'incorrect', 'wrong', 'sai', 'giả', 'không đúng']
        
        support_score = sum(1 for keyword in support_keywords if keyword in title_snippet)
        contradict_score = sum(1 for keyword in contradict_keywords if keyword in title_snippet)
        
        if contradict_score > support_score:
            contradict_count += 1
        elif support_score > contradict_score:
            support_count += 1
        else:
            # If unclear, assume slight support
            support_count += 0.5
    
    support_ratio = support_count / total_sources if total_sources > 0 else 0.5
    
    if support_ratio > 0.7:
        support_text = f"Sources strongly support the news: {support_count:.1f}/{total_sources} sources confirm"
    elif support_ratio > 0.4:
        support_text = f"Sources mixed: {support_count:.1f}/{total_sources} sources support, {contradict_count} contradict"
    else:
        support_text = f"Sources contradict the news: {contradict_count}/{total_sources} sources contradict"
    
    return support_ratio, support_text

def analyze_with_gemini(news_text, search_results, distilbert_prediction, distilbert_confidence):
    """Use Gemini AI to analyze the news and compare with our model results"""
    try:
        # Try to use the latest Gemini model available
        try:
            model = genai.GenerativeModel('gemini-2.0-flash-exp')
        except:
            try:
                model = genai.GenerativeModel('gemini-2.5-flash')
            except:
                try:
                    model = genai.GenerativeModel('gemini-1.5-pro')
                except:
                    model = genai.GenerativeModel('gemini-1.5-flash')
        
        # Format the search results for Gemini
        search_summary = ""
        if search_results:
            search_summary = "Kết quả tìm kiếm Google:\n"
            for i, result in enumerate(search_results[:5], 1):
                search_summary += f"{i}. {result['title']}\n   {result['snippet']}\n   Nguồn: {result['link']}\n\n"
        else:
            search_summary = "Không tìm thấy kết quả tìm kiếm Google cho tin tức này. Điều này có thể do API bị giới hạn hoặc tin tức quá mới/chưa được đăng tải."
        
        # Include our model results in the analysis
        distilbert_analysis = ""
        if distilbert_prediction:
            distilbert_analysis = f"Phân tích DistilBERT: Dự đoán '{distilbert_prediction}' với độ tin cậy {distilbert_confidence:.3f}"
        else:
            distilbert_analysis = "DistilBERT: Không thể phân tích"
        
        prompt = f"""
Hãy phân tích tin tức sau và đánh giá độ tin cậy của nó một cách đơn giản, dễ hiểu:

"{news_text}"

{search_summary}

{distilbert_analysis}

Hãy trả lời bằng tiếng Việt, ngắn gọn và dễ hiểu cho người dùng bình thường:

1. Tin tức này có vẻ THẬT hay GIẢ? (Chỉ trả lời THẬT hoặc GIẢ)
2. Tại sao bạn nghĩ vậy? (Giải thích ngắn gọn, dễ hiểu)
3. Người đọc nên làm gì? (Lời khuyên đơn giản)

Tránh dùng thuật ngữ kỹ thuật, hãy viết như đang nói chuyện với bạn bè.
"""
        
        print("Calling Gemini API...")
        print(f"DEBUG - News text being analyzed: {news_text}")
        print(f"DEBUG - Search results count: {len(search_results)}")
        if search_results:
            print(f"DEBUG - First search result title: {search_results[0].get('title', 'No title')}")
        
        # Use consistent settings to get reliable results
        generation_config = genai.types.GenerationConfig(
            temperature=0.1,  # Low temperature for more consistent results
            top_p=0.8,        # Focus on most likely tokens
            top_k=20,         # Limit vocabulary choices
            max_output_tokens=1000
        )
        response = model.generate_content(prompt, generation_config=generation_config)
        print("Gemini API response received successfully")
        return response.text
        
    except Exception as e:
        print(f"Gemini analysis error: {e}")
        print(f"Error type: {type(e).__name__}")
        
        # If we hit the API limit, provide a basic analysis
        if "429" in str(e) or "quota" in str(e).lower():
            print("Gemini API quota exceeded, providing fallback analysis...")
            fallback_analysis = f"""
**Phân tích cơ bản (do giới hạn API):**

🤖 **Kết quả AI:** {'Tin tức này có vẻ THẬT' if distilbert_prediction == 'REAL' else 'Tin tức này có vẻ GIẢ' if distilbert_prediction == 'FAKE' else 'Không thể xác định'}

📊 **Độ tin cậy:** {f"{distilbert_confidence:.0%}" if distilbert_confidence else 'Không có'}

🌐 **Nguồn tin:** {len(search_results) if search_results else 0} nguồn được tìm thấy

💡 **Khuyến nghị:** Hãy kiểm tra thêm từ các nguồn tin chính thống trước khi tin tưởng hoàn toàn.
"""
            return fallback_analysis
        
        # For other errors, see what models are available
        try:
            models = genai.list_models()
            print("Available models:")
            for model in models:
                if 'gemini' in model.name.lower():
                    print(f"  - {model.name}")
        except Exception as list_error:
            print(f"Could not list models: {list_error}")
        return f"Lỗi phân tích Gemini: {e}"

def calculate_combined_confidence(distilbert_prediction, distilbert_confidence, source_credibility, popularity_score, gemini_analysis, source_support=0.5):
    """Calculate combined confidence from all three tools"""
    
    # Base confidence from DistilBERT
    if distilbert_prediction == "REAL":
        base_confidence = distilbert_confidence
    else:
        base_confidence = 1 - distilbert_confidence
    
    # Adjust based on source credibility (stronger adjustments)
    if source_credibility > 0.7:
        credibility_adjustment = 0.2  # Increased from 0.1
    elif source_credibility > 0.4:
        credibility_adjustment = 0.05  # Small positive adjustment
    else:
        credibility_adjustment = -0.1
    
    # Adjust based on popularity
    if popularity_score > 0.7:
        popularity_adjustment = 0.1  # Increased from 0.05
    elif popularity_score > 0.4:
        popularity_adjustment = 0.0
    else:
        popularity_adjustment = -0.05
    
    # Adjust based on source support (whether sources support or contradict the news)
    if source_support > 0.7:
        support_adjustment = 0.15  # Sources strongly support
    elif source_support > 0.4:
        support_adjustment = 0.0   # Sources are neutral
    else:
        support_adjustment = -0.15  # Sources contradict
    
    # Adjust based on Gemini analysis (stronger adjustments)
    gemini_lower = gemini_analysis.lower()
    if "độ tin cậy cao" in gemini_lower or "tin cậy cao" in gemini_lower or "cao" in gemini_lower:
        gemini_adjustment = 0.2  # Increased from 0.1
    elif "độ tin cậy thấp" in gemini_lower or "tin cậy thấp" in gemini_lower or "thấp" in gemini_lower:
        gemini_adjustment = -0.2  # Increased from -0.1
    else:
        gemini_adjustment = 0.0
    
    # Special case: If DistilBERT confidence is very low but sources and Gemini agree it's real
    if (distilbert_confidence < 0.6 and 
        source_credibility > 0.6 and 
        ("cao" in gemini_lower or "chính xác" in gemini_lower or "đáng tin cậy" in gemini_lower) and
        not ("thấp" in gemini_lower or "giả" in gemini_lower or "fake" in gemini_lower)):
        # Override with higher confidence ONLY if Gemini says it's real
        base_confidence = 0.8
        print("Overriding low DistilBERT confidence due to strong source and Gemini agreement for REAL")
    
    # Special case: If DistilBERT and Gemini both say FAKE, respect that
    elif (distilbert_prediction == "FAKE" and 
          ("thấp" in gemini_lower or "giả" in gemini_lower or "fake" in gemini_lower)):
        # Override with low confidence for FAKE
        base_confidence = 0.2
        print("Overriding confidence due to DistilBERT and Gemini agreement for FAKE")
    
    # Calculate final confidence
    final_confidence = base_confidence + credibility_adjustment + popularity_adjustment + gemini_adjustment + support_adjustment
    final_confidence = max(0.05, min(0.95, final_confidence))
    
    return final_confidence

def analyze_news(news_text):
    """Main analysis function combining all three tools"""
    try:
        if not news_text.strip():
            empty_message = """
<div style="font-family: 'Segoe UI', Arial, sans-serif; line-height: 1.6; color: #333;">

## 📝 **HƯỚNG DẪN SỬ DỤNG**

<div style="background: linear-gradient(135deg, #74b9ff 0%, #0984e3 100%); color: white; padding: 20px; border-radius: 10px; margin: 20px 0; text-align: center;">
<h2 style="margin: 0; font-size: 24px;">💡 Vui lòng nhập tin tức</h2>
<p style="margin: 10px 0 0 0; font-size: 16px; opacity: 0.9;">Để bắt đầu phân tích</p>
</div>

<div style="background: #f8f9fa; padding: 15px; border-radius: 8px; border-left: 4px solid #17a2b8; margin: 10px 0;">
<p><strong>Hướng dẫn:</strong></p>
<ul>
<li>Nhập tin tức tiếng Việt cần kiểm tra vào ô trên</li>
<li>Nhấn nút "Phân tích với AI nâng cao"</li>
<li>Chờ hệ thống phân tích (có thể mất 10-30 giây)</li>
<li>Xem kết quả phân tích chi tiết</li>
</ul>
</div>

</div>
"""
            return gr.update(value=empty_message, visible=True), "Độ chắc chắn là tin thật: 0%", "Độ chắc chắn là tin giả: 0%", gr.update(visible=False)
        
        print(f"Analyzing: {news_text[:50]}...")
        
        # Step 1: Search Google for related information
        print("1. Running Google Search...")
        try:
            search_results = google_search(news_text)
        except Exception as e:
            print(f"Google Search error: {e}")
            search_results = []
        
        # Step 2: Run our trained model
        print("2. Running DistilBERT analysis...")
        try:
            distilbert_prediction, distilbert_confidence, real_score, fake_score = predict_with_distilbert(news_text)
        except Exception as e:
            print(f"DistilBERT analysis error: {e}")
            distilbert_prediction, distilbert_confidence, real_score, fake_score = None, None, None, None
        
        # Step 3: Check the sources we found
        print("3. Analyzing sources and popularity...")
        try:
            source_credibility, popularity_score, credibility_text = analyze_sources(search_results)
            source_support, support_text = analyze_source_support(news_text, search_results)
        except Exception as e:
            print(f"Source analysis error: {e}")
            source_credibility, popularity_score, credibility_text = 0.5, 0.2, "Lỗi phân tích nguồn"
            source_support, support_text = 0.5, "Lỗi phân tích hỗ trợ nguồn"
        
        # Step 4: Get Gemini AI analysis
        print("4. Running Gemini analysis...")
        try:
            gemini_analysis = analyze_with_gemini(news_text, search_results, distilbert_prediction, distilbert_confidence)
        except Exception as e:
            print(f"Gemini analysis error: {e}")
            gemini_analysis = f"Lỗi phân tích Gemini: {str(e)}"
        
        # Step 5: Combine everything into final result
        print("5. Calculating combined confidence...")
        print(f"   DistilBERT: {distilbert_prediction} ({distilbert_confidence:.3f})")
        print(f"   Source credibility: {source_credibility:.3f}")
        print(f"   Source support: {source_support:.3f}")
        print(f"   Popularity: {popularity_score:.3f}")
        try:
            combined_confidence = calculate_combined_confidence(
                distilbert_prediction, distilbert_confidence, 
                source_credibility, popularity_score, gemini_analysis, source_support
            )
            print(f"   Final combined confidence: {combined_confidence:.3f}")
        except Exception as e:
            print(f"Confidence calculation error: {e}")
            combined_confidence = 0.5  # Default to neutral
        
        # Step 6: Format the final results
        real_confidence = combined_confidence
        fake_confidence = 1 - combined_confidence
    
        # Build the detailed report with better formatting
        prediction_emoji = "✅" if distilbert_prediction == "REAL" else "❌" if distilbert_prediction == "FAKE" else "❓"
        confidence_level = "Cao" if combined_confidence > 0.7 else "Trung bình" if combined_confidence > 0.4 else "Thấp"
        confidence_emoji = "🟢" if combined_confidence > 0.7 else "🟡" if combined_confidence > 0.4 else "🔴"
        
        # Convert technical metrics to user-friendly Vietnamese
        source_quality = "Tốt" if source_credibility > 0.7 else "Trung bình" if source_credibility > 0.4 else "Kém"
        source_count_text = f"{len(search_results)} nguồn tin" if len(search_results) > 0 else "Không tìm thấy nguồn"
        
        # Simplify credibility text
        if "High credibility" in credibility_text:
            credibility_summary = f"✅ Nguồn tin đáng tin cậy"
        elif "Medium credibility" in credibility_text:
            credibility_summary = f"⚠️ Nguồn tin trung bình"
        else:
            credibility_summary = f"❌ Nguồn tin kém tin cậy"
        
        # Simplify support text
        if "strongly support" in support_text.lower():
            support_summary = "✅ Các nguồn ủng hộ tin tức này"
        elif "contradict" in support_text.lower():
            support_summary = "❌ Các nguồn phản bác tin tức này"
        else:
            support_summary = "⚠️ Các nguồn có ý kiến trái chiều"
        
        detailed_analysis = f"""
<div style="font-family: 'Segoe UI', Arial, sans-serif; line-height: 1.6; color: #333;">

## 🔍 **KẾT QUẢ PHÂN TÍCH TIN TỨC**

<div style="background: linear-gradient(135deg, #667eea 0%, #764ba2 100%); color: white; padding: 20px; border-radius: 10px; margin: 20px 0; text-align: center;">
<h2 style="margin: 0; font-size: 24px;">{prediction_emoji} {'TIN THẬT' if distilbert_prediction == 'REAL' else 'TIN GIẢ' if distilbert_prediction == 'FAKE' else 'KHÔNG XÁC ĐỊNH'}</h2>
<p style="margin: 10px 0 0 0; font-size: 18px; opacity: 0.9;">{confidence_emoji} Độ tin cậy: {confidence_level} ({combined_confidence:.0%})</p>
</div>

### 🤖 **Phân tích bằng AI**
<div style="background: #f8f9fa; padding: 15px; border-radius: 8px; border-left: 4px solid #007bff; margin: 10px 0;">
<p><strong>Kết quả:</strong> {prediction_emoji} {'Tin tức này có vẻ THẬT' if distilbert_prediction == 'REAL' else 'Tin tức này có vẻ GIẢ' if distilbert_prediction == 'FAKE' else 'Không thể xác định'}</p>
<p><strong>Độ chắc chắn:</strong> {f"{distilbert_confidence:.0%}" if distilbert_confidence else 'Không có'} - {'Rất cao' if distilbert_confidence and distilbert_confidence > 0.8 else 'Cao' if distilbert_confidence and distilbert_confidence > 0.6 else 'Trung bình' if distilbert_confidence and distilbert_confidence > 0.4 else 'Thấp'}</p>
</div>

### 🌐 **Kiểm tra nguồn tin**
<div style="background: #f8f9fa; padding: 15px; border-radius: 8px; border-left: 4px solid #28a745; margin: 10px 0;">
<p><strong>Tìm thấy:</strong> {source_count_text}</p>
<p><strong>Chất lượng nguồn:</strong> {source_quality} ({source_credibility:.0%})</p>
<p><strong>Đánh giá:</strong> {credibility_summary}</p>
<p><strong>Hỗ trợ:</strong> {support_summary}</p>
</div>

### 🧠 **Phân tích thông minh**
<div style="background: #f8f9fa; padding: 15px; border-radius: 8px; border-left: 4px solid #ffc107; margin: 10px 0;">
{gemini_analysis}
</div>

### 📊 **KẾT LUẬN CUỐI CÙNG**
<div style="background: linear-gradient(135deg, #f093fb 0%, #f5576c 100%); color: white; padding: 15px; border-radius: 8px; margin: 20px 0;">
<p style="margin: 0; font-size: 16px;"><strong>Tin tức này có khả năng {'THẬT' if real_confidence > fake_confidence else 'GIẢ'} với độ tin cậy {max(real_confidence, fake_confidence):.0%}</strong></p>
<p style="margin: 5px 0 0 0; font-size: 14px; opacity: 0.9;">Dựa trên phân tích AI, kiểm tra nguồn tin và đánh giá thông minh</p>
</div>

</div>
"""
    
        return gr.update(value=detailed_analysis, visible=True), f"Độ chắc chắn là tin thật: {real_confidence:.1%}", f"Độ chắc chắn là tin giả: {fake_confidence:.1%}", gr.update(visible=True)
    
    except Exception as e:
        error_message = f"""
<div style="font-family: 'Segoe UI', Arial, sans-serif; line-height: 1.6; color: #333;">

## ❌ **LỖI PHÂN TÍCH**

<div style="background: linear-gradient(135deg, #ff6b6b 0%, #ee5a24 100%); color: white; padding: 20px; border-radius: 10px; margin: 20px 0; text-align: center;">
<h2 style="margin: 0; font-size: 24px;">⚠️ Có lỗi xảy ra</h2>
<p style="margin: 10px 0 0 0; font-size: 16px; opacity: 0.9;">Vui lòng thử lại sau</p>
</div>

<div style="background: #f8f9fa; padding: 15px; border-radius: 8px; border-left: 4px solid #dc3545; margin: 10px 0;">
<p><strong>Chi tiết lỗi:</strong> {str(e)}</p>
<p><strong>Gợi ý:</strong> Kiểm tra kết nối internet và thử lại</p>
</div>

</div>
"""
        print(f"Analysis error: {e}")
        return gr.update(value=error_message, visible=True), "Độ chắc chắn là tin thật: 0%", "Độ chắc chắn là tin giả: 0%", gr.update(visible=True)

# --- GRADIO INTERFACE ---
def create_interface():
    with gr.Blocks(title="Vietnamese Fake News Detection System", theme=gr.themes.Soft()) as interface:
        gr.Markdown("""
        <div style="text-align: center; padding: 20px;">
        <h1 style="color: #2c3e50; margin-bottom: 10px;">🔍 Vietnamese Fake News Detection System</h1>
        <p style="color: #7f8c8d; font-size: 16px; margin-bottom: 30px;">Powered by Google Search + Gemini AI + DistilBERT</p>
        
        <div style="background: linear-gradient(135deg, #667eea 0%, #764ba2 100%); color: white; padding: 20px; border-radius: 15px; margin: 20px 0;">
        <h3 style="margin: 0 0 15px 0;">🛡️ Hệ thống phát hiện tin giả tiếng Việt</h3>
        <div style="display: flex; justify-content: space-around; flex-wrap: wrap; gap: 15px;">
        <div style="text-align: center;">
        <div style="font-size: 24px; margin-bottom: 5px;">🌐</div>
        <strong>Google Search</strong><br>
        <small>Tìm kiếm thông tin thực tế</small>
        </div>
        <div style="text-align: center;">
        <div style="font-size: 24px; margin-bottom: 5px;">🧠</div>
        <strong>Gemini AI</strong><br>
        <small>Phân tích thông minh</small>
        </div>
        <div style="text-align: center;">
        <div style="font-size: 24px; margin-bottom: 5px;">🤖</div>
        <strong>DistilBERT</strong><br>
        <small>AI chuyên tiếng Việt</small>
        </div>
        </div>
        </div>
        
        <div style="background: #f8f9fa; padding: 15px; border-radius: 10px; border-left: 4px solid #17a2b8; margin: 20px 0;">
        <p style="margin: 0; color: #495057;"><strong>💡 Lưu ý:</strong> Kết quả có thể thay đổi nhẹ giữa các lần phân tích do tính chất AI của Gemini, nhưng độ chính xác tổng thể vẫn được đảm bảo.</p>
        </div>
        </div>
        """)
        
        with gr.Row():
            with gr.Column(scale=2):
                gr.Markdown("### 📝 Nhập tin tức cần kiểm tra")
                news_input = gr.Textbox(
                    placeholder="Nhập tin tức tiếng Việt cần kiểm tra...",
                    lines=4,
                    show_label=False
                )
                
                analyze_btn = gr.Button("🔍 Phân tích với AI nâng cao", variant="primary", size="lg")
            
            with gr.Column(scale=1, visible=False) as results_column:
                gr.Markdown("### 📊 Kết quả phân tích")
                real_confidence = gr.Label(value="Độ chắc chắn là tin thật: 0%")
                fake_confidence = gr.Label(value="Độ chắc chắn là tin giả: 0%")
        
        detailed_analysis = gr.Markdown("### 📋 Phân tích chi tiết sẽ hiển thị ở đây...", visible=False)
        
        # Event handlers
        analyze_btn.click(
            fn=analyze_news,
            inputs=[news_input],
            outputs=[detailed_analysis, real_confidence, fake_confidence, results_column]
        )
        

    return interface

def test_google_search():
    """Test Google Search API functionality"""
    print("Testing Google Search API...")
    print("=" * 50)
    
    # Test queries
    test_queries = [
        "Argentina World Cup 2022",
        "Vietnam COVID-19 news",
        "Tin tức Việt Nam"
    ]
    
    results_found = 0
    
    for i, query in enumerate(test_queries, 1):
        print(f"\nTest {i}: '{query}'")
        print("-" * 30)
        
        try:
            results = google_search(query)
            print(f"Results: {len(results)} found")
            
            if results:
                results_found += 1
                print(f"First result: {results[0]['title'][:50]}...")
                print(f"   Link: {results[0]['link']}")
            else:
                print("No results found")
                
        except Exception as e:
            print(f"Error: {e}")
    
    print(f"\nTest Summary: {results_found}/{len(test_queries)} tests passed")
    
    if results_found == 0:
        print("\nGoogle Search is not working!")
        print("Possible solutions:")
        print("   1. Check API quota in Google Cloud Console")
        print("   2. Verify API keys are correct")
        print("   3. Ensure Custom Search API is enabled")
        print("   4. Check Search Engine ID is valid")
    elif results_found < len(test_queries):
        print("\nGoogle Search partially working")
        print("Some queries work, others don't - check query formatting")
    else:
        print("\nGoogle Search is working perfectly!")
    
    return results_found > 0

def test_complete_system():
    """Test the complete fake news detection system"""
    print("Testing Complete Vietnamese Fake News Detection System")
    print("=" * 60)
    
    # Test cases
    test_cases = [
        "Argentina vô địch World Cup 2022",
        "Hôm nay trời mưa ở Hà Nội",
        "COVID-19 đã được chữa khỏi hoàn toàn"
    ]
    
    for i, test_text in enumerate(test_cases, 1):
        print(f"\nTest Case {i}: '{test_text}'")
        print("-" * 40)
        
        try:
            result = analyze_news(test_text)
            print("Analysis completed successfully")
            print(f"Result type: {type(result)}")
        except Exception as e:
            print(f"Analysis failed: {e}")

# --- LAUNCH APP ---
if __name__ == "__main__":
    print("Starting Vietnamese Fake News Detection System...")
    print("Tools integrated: Google Search + Gemini AI + DistilBERT")
    
    # Uncomment the line below to run tests first
    # test_google_search()
    
    interface = create_interface()
    interface.launch(
        server_name="0.0.0.0",
        server_port=7860,  # Standard port for Hugging Face Spaces
        share=False,  # Not needed for Hugging Face Spaces
        show_error=True
    )