🔍 AI News Verification System

An LLM + RAG-based evidence verification system that verifies news URLs and viral text claims against real-time reporting from global and regional news sources without relying on a traditional supervised fake-news classifier.

---

🎯 Core Product Objective

When encountering a suspicious claim on social media or browsing the web, users should not have to manually visit multiple news portals one by one.

VerifyNews automates this multi-source investigation:

1. Ingests a news article URL or pasted text claim.
2. Extracts the core verifiable claim(s) and named entities using Groq LLM.
3. Formulates targeted search queries across multiple angles.
4. Retrieves real-time news articles from NewsAPI and Google News (SerpAPI).
5. Deduplicates cross-publisher syndicated wire stories.
6. Indexes & Semantically Searches evidence using FastEmbed (BAAI/bge-small-en-v1.5) and in-memory cosine similarity search.
7. Evaluates stance (SUPPORT, CONTRADICT, NEUTRAL) and context for each retrieved source using Groq LLM.
8. Synthesizes a transparent verdict: REAL, FALSE, MISLEADING, or UNVERIFIED with evidence snippets and confidence breakdown.

---

🏗️ Architecture

USER INPUT (URL or Text)
       │
       ▼
 FastAPI Backend (/api/v1/verify)
       │
       ├─► [If URL] Article Extractor (newspaper3k + bs4)
       │
       ▼
 Groq LLM — Claim Extraction (Primary claim, Sub-claims, Entities)
       │
       ▼
 Groq LLM — Search Query Generation (3-5 targeted queries)
       │
       ├─────────────────────────┬─────────────────────────┐
       ▼                         ▼                         ▼
   NewsAPI Search         SerpAPI Google News        Fallback Store
       │                         │                         │
       └─────────────────────────┼─────────────────────────┘
                                 ▼
                     Source Normalization & Wire Deduplication
                                 ▼
                      FastEmbed (384-dim Dense Vectors)
                                  ▼
                      In-Memory Semantic Vector Index
                                  ▼
                      Semantic Evidence Retrieval (Top-K)
                                 ▼
                     Groq LLM — Stance & Nuance Analysis (SUPPORT / CONTRADICT / NEUTRAL)
                                 ▼
                     Groq LLM — Final Verdict & Transparent Synthesis
                                 │
                                 ▼
                     REAL / FALSE / MISLEADING / UNVERIFIED
                                 │
                                 ▼
                     Interactive Single-Page UI

---

🏷️ Verdict Definitions

• REAL: Multiple authoritative retrieved sources independently confirm and support the central claim.
• FALSE: Authoritative retrieved evidence directly contradicts or refutes the central claim.
• MISLEADING: The claim contains partial truth but has missing context, exaggerates numbers, or presents outdated events as current.
• UNVERIFIED: Retrieved evidence is insufficient, inconclusive, or completely neutral.

---

🚀 Quick Start

1. Prerequisites

• Python 3.10+
• Free Groq API Key
• Free NewsAPI Key
• (Optional) SerpAPI Key

2. Installation

git clone https://github.com/jaiyan-th/Fake-News-Detecter.git
cd Fake-News-Detecter

# Install dependencies
pip install -r requirements.txt

3. Environment Configuration

Create a .env file in the project root:

GROQ_API_KEY=your-groq-api-key
NEWS_API_KEY=your-news-api-key
SERPAPI_KEY=your-serpapi-key-optional

4. Start the Application

# Start FastAPI backend & UI
uvicorn backend.main:app --reload --port 8000

Open your browser at:

• Interactive UI: http://localhost:8000
• Swagger / OpenAPI Documentation: http://localhost:8000/docs
• System Health Diagnostics: http://localhost:8000/api/v1/health

---

📡 API Reference

POST /api/v1/verify

Verify a news URL or text claim.

Request Body:

{
  "url": "https://reuters.com/world/india/example-article",
  "text": null
}

Response (200 OK):

{
  "verdict": "REAL",
  "confidence": 88,
  "confidence_label": "Verification Confidence",
  "claim": {
    "primary_claim": "India launched new solar energy initiative.",
    "secondary_claims": [],
    "entities": ["India", "Solar Mission"],
    "timeframe": "Recent"
  },
  "summary": "Multiple reputable sources confirm the government announcement.",
  "explanation": "Reuters and The Hindu both report that the cabinet approved the solar outlay today.",
  "evidence_summary": {
    "supporting": 2,
    "contradicting": 0,
    "neutral": 1,
    "total_sources_evaluated": 3
  },
  "sources": [
    {
      "source_name": "Reuters",
      "domain": "reuters.com",
      "title": "Cabinet Clears Solar Plan",
      "url": "https://reuters.com/...",
      "stance": "SUPPORT",
      "relevance_score": 0.94,
      "evidence_snippet": "The government approved the funding.",
      "credibility_tier": "WIRE_AND_PRIMARY_AGENCY"
    }
  ],
  "source_agreement_percentage": 100.0,
  "limitations": [
    "Analysis based on English-language articles available at query time."
  ],
  "pipeline_stages": [
    { "stage": "article_extraction", "status": "COMPLETED", "duration_ms": 350 },
    { "stage": "claim_extraction", "status": "COMPLETED", "duration_ms": 480 },
    { "stage": "multi_source_search", "status": "COMPLETED", "duration_ms": 1100 },
    { "stage": "semantic_vector_retrieval", "status": "COMPLETED", "duration_ms": 120 },
    { "stage": "verdict_synthesis", "status": "COMPLETED", "duration_ms": 650 }
  ],
  "processing_time_ms": 2700
}

---

🧪 Running Tests

python -m pytest backend/tests/ -v

---

📄 License

MIT License. Created for AI-powered news verification and claim investigation.