README.md

metadata

title: AgenticClaimVerificationSystem
emoji: 🔥
colorFrom: purple
colorTo: red
sdk: gradio
sdk_version: 5.33.1
app_file: app.py
pinned: false
short_description: Multi Agentic Claim Verification System
tags:
  - agent-demo-track

🔍 Multi-Agent Claim Verification System using Kognie API

An intelligent, multi-agent system designed to verify claims using diverse AI models and real-time web research. This system combines the power of multiple language models with web search capabilities to provide comprehensive fact-checking and evidence analysis.

Visit : https://kognie.com/api to create a Kognie API key to gain access to multiple LLMs with one single account.

📽️ Demo videos

Showcasing the app and explaining the architecture under the hood of the claim verification system. URL : https://drive.google.com/file/d/10L3T1Hvie9D6XqMI7H80hCp98iVuP3Nm/view?usp=sharing

🎯 Purpose

In an era of information overload and misinformation, this system serves as a robust fact-checking tool that:

• Verifies claims using multiple AI perspectives
• Gathers real-time evidence from web sources
• Provides balanced analysis with supporting and contradicting evidence
• Makes informed decisions based on comprehensive data analysis
• Presents results in an intuitive, interactive web interface

🏗️ System Architecture

The system employs a hierarchical multi-agent architecture with specialized roles:

🤖 Agent Specifications

1. MultiLLM Verifier Agent

• Model: Claude-3.5-Sonnet (Anthropic)

• Role: Cross-model evidence analysis

• Responsibilities:

  • Coordinates multiple LLM perspectives
  • Runs parallel analysis across different AI models
  • Provides diverse viewpoints on claims

• Internal process: The system leverages three distinct AI models for comprehensive analysis:

  Model	Provider	Strengths
  GPT-4o-mini	Kognie API	Fast reasoning, general knowledge
  Gemini-2.0-Flash	Kognie API	Multimodal capabilities, recent training
  Open-Mistral-Nemo	Kognie API	European perspective, specialized domains

  Parallel Processing Benefits

  • Diverse Perspectives: Each model brings unique training and biases
  • Cross-Validation: Multiple viewpoints reduce single-model limitations
  • Speed: Asynchronous processing ensures rapid results
  • Robustness: System continues functioning even if one model fails

2. Web Evidence Retriever Agent

• Model: Claude-3.5-Sonnet (Anthropic)

• Role: Real-time information gathering

• Responsibilities:

  • Searches current web sources
  • Retrieves up-to-date information
  • Provides context-aware evidence

• Internal process:

  Real-Time Evidence Gathering

  • Bing Search API integration for current information
  • News source prioritization for recent developments
  • Automated query generation based on claim analysis
  • Evidence categorization (supporting vs. contradicting)

  Search Strategy

  1. Query Optimization: Transforms claims into effective search terms
  2. Source Diversification: Gathers information from multiple web sources
  3. Recency Prioritization: Focuses on current and relevant information
  4. Result Synthesis: Analyzes and structures findings

3. Boss Agent (Coordinator)

• Model: GPT-4o (OpenAI)
• Role: Final decision maker and coordinator
• Responsibilities:
  • Orchestrates other agents
  • Synthesizes evidence from multiple sources
  • Makes final verification decisions
  • Formats results in HTML for presentation

💻 User Interface

Interactive Web Interface (Gradio)

• Chat-based interaction for natural claim submission
• Real-time processing with progress indicators
• Collapsible analysis sections for detailed evidence review
• Color-coded results (Green for TRUE, Red for FALSE)
• Responsive design for various devices

Key Features

• Instant verification results
• Detailed evidence breakdown from each agent
• Interactive expandable sections for in-depth analysis
• Clean, professional presentation of complex data

🚀 Getting Started

Prerequisites

pip install gradio llama-index python-dotenv asyncio

Environment Variables

Create a .env file with the following:

KOGNIE_BASE_URL=your_kognie_base_url
KOGNIE_API_KEY=your_kognie_api_key
BING_SUBSCRIPTION_KEY=your_bing_api_key
BING_SEARCH_URL=your_bing_search_url
ANTHROPIC_API_KEY=your_anthropic_api_key
OPENAI_API_KEY=your_openai_api_key
MISTRAL_API_KEY=your_mistral_api_key

Running the Application

gradio app.py

The system will launch a web interface accessible through your browser.

🎯 Use Cases

Perfect For:

• Fact-checking news claims
• Academic research verification
• Social media post validation
• Business claim analysis
• Educational fact verification
• Journalism and reporting

Example Claims:

• "Company X reported record profits in Q4 2024"
• "New scientific study proves Y causes Z"
• "Political candidate made statement about policy"
• "Sports team won championship in specific year"

🔮 Technical Advantages

1. Asynchronous Processing

• Non-blocking operations for faster results
• Concurrent agent execution
• Responsive user interface

2. Error Resilience

• Graceful handling of API failures
• Fallback mechanisms for each component
• Comprehensive error logging

3. Scalable Architecture

• Easy addition of new AI models
• Modular agent design
• Configurable processing parameters

4. Evidence Transparency

• Complete audit trail of analysis
• Source attribution for all evidence
• Detailed reasoning for decisions

🛡️ Quality Assurance

Multi-Layer Verification

1. Cross-Model Validation: Multiple AI perspectives
2. Real-Time Research: Current information priority
3. Evidence Weighting: Web sources prioritized for recent events
4. Transparent Reasoning: Complete decision audit trail

Bias Mitigation

• Model Diversity: Different training approaches and datasets
• Source Variety: Multiple web sources and perspectives
• Temporal Awareness: Prioritizes recent information
• Evidence Balance: Seeks both supporting and contradicting evidence

🔧 Customization Options

The system is designed for easy customization:

• Add new AI models to the MultiLLM verifier
• Integrate additional search engines beyond Bing
• Modify decision-making logic in the Boss Agent
• Customize UI themes and presentation styles
• Adjust evidence weighting algorithms