VeriFact

AI-powered fact-checking system with evidence-backed explanations

alt text

A fact-checking system that verifies user claims with evidence-backed explanations using web searches and AI, delivering clear, transparent verdicts through a simple interface.

Key Features

  • Explainable Verdicts: Generates clear, evidence-based explanations with references to credible web sources
  • Machine Learning Integration: Uses Sentence Transformers for semantic similarity and BART-MNLI for natural language inference (NLI)
  • Text Polishing: Employs Pegasus-XSUM to rephrase explanations for clarity and readability
  • Efficient Caching: Stores results in PostgreSQL to avoid redundant processing
  • Multi-Channel Access: Web frontend (Vue.js), CLI, and Telegram bot integration
  • Scalable Backend: FastAPI with async support for handling multiple concurrent requests
  • Web Search Integration: Leverages Google Custom Search to retrieve relevant sources
  • Confidence Scoring: Provides numerical confidence in claim assessment

How It Works

  1. Claim Input: Users submit claims via CLI, web interface, or Telegram bot
  2. Cache Check: System queries PostgreSQL for cached results using normalized claim
  3. Web Search: Google Custom Search API retrieves up to 10 relevant URLs
  4. Heuristic Analysis: Scores search results using keyword-based heuristics
  5. Deep ML Analysis: Uses Sentence Transformers and BART-MNLI to classify evidence as ENTAILMENT, CONTRADICTION, or NEUTRAL
  6. Verdict Fusion: Combines heuristic and ML results for final verdict
  7. Explanation Generation: Constructs factual explanation with supporting evidence
  8. Text Polishing: Rephrases explanation using Pegasus-XSUM for clarity
  9. Response & Caching: Returns verdict with confidence score and stores result in PostgreSQL

Technologies Used

  • Backend: FastAPI (Python)
  • ML Models: PyTorch, Sentence Transformers, BART-MNLI, Pegasus-XSUM
  • Frontend: Vue.js 3 + Vite
  • Database: PostgreSQL
  • Web Scraping: Trafilatura
  • Search: Google Custom Search API
  • Additional: Telegram bot integration, CLI interface

Verdict Classifications

  • Likely True: Strong evidence with supporting sources
  • Likely False: Contradicting evidence from credible sources
  • Mixed/Uncertain: Conflicting evidence or insufficient data

Real-World Impact

  • HackIndore Achievement: Ranked 7th out of 200 teams at HackIndore hackathon
  • Paid Contract: The project’s success led to a paid development contract, validating the practical viability of the fact-checking system
  • Production-Ready: Demonstrated capability to combat misinformation through an accessible, evidence-backed verification platform
View on GitHub