README.md

metadata

title: VerifAI News Verification API
emoji: 🔍
colorFrom: purple
colorTo: blue
sdk: docker
app_port: 7860
pinned: false
license: mit

VerifAI - Fake News Detection API

Multi-model fake news detection system for both text and images with intelligent API management and diverse global news sources.

Version: 3.0.0
Last Updated: January 2026

✨ Features

Core Detection

• Text Detection: BERT + LLM (Gemini 2.0/Groq) + Web Verification
• Image Detection: OCR (EasyOCR) + Gemini Vision + Manipulation Analysis
• Real-time API: FastAPI with CORS support

Intelligent API Management

• 6 Gemini API Keys with smart round-robin rotation
• Groq Llama 3.1 70B Fallback when Gemini is exhausted
• Response Caching with 1-hour TTL (40-60% API reduction)
• Rate Limit Detection with 60-second cooldowns

Diverse News Sources

• 100+ Trusted Domains from all continents
• Unique Domain Per Source (no duplicate domains)
• Wire Services: Reuters, AP, AFP
• International: BBC, CNN, DW, Al Jazeera, NDTV, SCMP
• Regional coverage: US, UK, Europe, Asia, Africa, Latin America, Oceania

---

📡 API Endpoints

Health Check

GET /

Returns API status including cache stats and key health.

API Key Status

GET /api-status

Returns detailed status of all Gemini keys and Groq fallback.

Text Verification

POST /check
Content-Type: application/json

{"text": "Your news text to verify"}

Response:

{
    "credibility": 85.0,
    "verdict": "VERIFIED",
    "bert_score": 0.92,
    "llm_score": 0.78,
    "web_score": 1.0,
    "sources": [
        {
            "title": "BBC News - Article Title",
            "url": "https://bbc.com/...",
            "trusted": true,
            "domain": "bbc.com"
        }
    ]
}

Image Verification (Base64)

POST /check-image
Content-Type: application/json

{"image": "data:image/jpeg;base64,..."}

Response:

{
    "credibility": 75,
    "verdict": "LIKELY REAL",
    "alert": "✅ Content appears mostly credible",
    "extracted_text": "Text extracted from image...",
    "bert_score": 0.85,
    "llm_score": 0.70,
    "web_score": 0.8,
    "image_manipulation_score": 0.2,
    "concerns": "Minor artifacts detected",
    "sources": [...]
}

Image Verification (File Upload)

POST /check-image-upload
Content-Type: multipart/form-data

image: <file>

Text Extraction (OCR only)

POST /extract-text
Content-Type: application/json

{"image": "data:image/jpeg;base64,..."}

---

🔧 Environment Variables

Set these in Space Settings → Repository secrets:

Variable	Description	Required
GEMINI_API_KEY	Additional Gemini API key	Optional
GROQ_API_KEY	Groq API key for fallback	Pre-configured

Note: 6 Gemini keys are already hardcoded in app.py. Environment variables are optional extras.

---

📊 Credibility Scoring

Score Range	Verdict	Description
80-100%	✅ VERIFIED / REAL	Highly credible, matched trusted sources
60-79%	⚠️ LIKELY REAL	Mostly credible, some uncertainty
40-59%	🤔 UNCERTAIN	Inconclusive, requires manual review
20-39%	⚠️ LIKELY FAKE	Suspicious patterns detected
0-19%	🚨 FAKE	Strong indicators of misinformation

---

🌐 Trusted News Sources

Wire Services: Reuters, AP, AFP, UPI

By Region:

Region	Example Sources
US	NYT, WaPo, WSJ, CNN, NBC, Bloomberg, Forbes
UK	BBC, Guardian, Telegraph, FT, Economist
Europe	DW (Germany), France24, El País, Corriere
Middle East	Al Jazeera, Arab News, Times of Israel
Asia	NDTV, Hindu, SCMP, Japan Times, Korea Herald
Africa	News24, Daily Maverick, AllAfrica
Latin America	Folha, Infobae, El Universal
Oceania	ABC Australia, NZ Herald

---

🚀 Deployment

Hugging Face Spaces

1. Create a new Space with Docker SDK
2. Upload contents of huggingface/ folder
3. Space will auto-build and deploy

Local Development

cd huggingface
pip install -r requirements.txt
python -m uvicorn app:app --host 127.0.0.1 --port 7860 --reload

---

📝 Model Information

Component	Model	Purpose
BERT	jy46604790/Fake-News-BERT-Detect	Binary classification
LLM	Gemini 2.0 Flash / Groq Llama 3.1 70B	Fact analysis
Vision	Gemini Vision	Image manipulation detection
OCR	EasyOCR	Text extraction from images
Search	DuckDuckGo	Web verification

---

📈 Performance

• Startup Time: ~30-60s (model loading)
• Text Analysis: 3-8s per request
• Image Analysis: 5-12s per request
• Cache Hit Rate: 40-60% (reduces API calls)
• Uptime: 99.9% (Gemini + Groq redundancy)

---

📄 License

MIT License