Veritas

An intelligent multi-agent research platform that performs automated research, claim extraction, contradiction detection, verification, and confidence scoring using a hybrid architecture combining LLM agents, semantic retrieval, and local ML models.

---

🚀 Overview

This project is an AI-powered research assistant system that:

• Breaks complex questions into research tasks
• Searches and analyzes sources
• Extracts structured factual claims
• Detects contradictions across sources
• Verifies evidence using hybrid ML + LLM reasoning
• Assigns confidence scores
• Generates explainable research reports

The system combines:

• Multi-agent orchestration
• Retrieval-augmented reasoning
• Semantic vector search
• Transformer-based stance classification
• Structured knowledge storage

---

🧠 Key Features

✅ Multi-Agent Architecture

Specialized agents collaborate to solve research tasks:

• Planner Agent

  • Breaks queries into sub-questions

• Research Agent

  • Retrieves and summarizes evidence

• Claim Extraction Agent

  • Converts text into structured factual claims

• Verification Agent

  • Evaluates evidence and claim validity

• Writer Agent

  • Synthesizes structured final reports

---

🔎 Semantic Knowledge Layer

Claims are stored in ChromaDB with embeddings for:

• Semantic retrieval
• Topic clustering
• Contradiction analysis
• Persistent research memory

---

⚖️ Contradiction Detection

The system identifies:

• Supporting claims
• Opposing claims
• Uncertain claims

using:

• Embedding similarity
• Transformer-based NLI models
• LLM reasoning

---

🧪 Hybrid ML + LLM Reasoning

The platform combines:

• Local transformer inference
• Embedding similarity scoring
• LLM-based nuanced reasoning

instead of relying entirely on LLM prompts.

---

📊 Explainable AI Pipeline

The system exposes:

• Agent execution traces
• Evidence sources
• Confidence reasoning
• Contradiction analysis

to improve transparency and interpretability.

---

🏗️ System Architecture

Frontend (React / Next.js)
        ↓
FastAPI Backend
        ↓
Agent Orchestrator
        ↓
Agents:
  ├── Planner Agent
  ├── Research Agent
  ├── Claim Extraction Agent
  ├── Verification Agent
  └── Writer Agent
        ↓
ML Intelligence Layer
  ├── Embedding Similarity
  ├── Stance Classification
  └── Confidence Scoring
        ↓
ChromaDB Vector Store
        ↓
External Search + Scraping

---

⚙️ Tech Stack

Backend

• Python
• FastAPI
• Google Agent Development Kit (ADK)

Frontend

• React / Next.js

AI / ML

• Gemini API
• Hugging Face Transformers
• PyTorch
• Sentence Transformers

Vector Storage

• ChromaDB

Search / Retrieval

• Tavily API
• BeautifulSoup
• newspaper3k

Deployment

• Docker
• Docker Compose
• Nginx
• AWS EC2
• GitHub Actions CI/CD

---

🧩 Core Pipeline

1. Planning

The Planner Agent decomposes the user query into focused research tasks.

Example:

{
  "tasks": [
    {
      "id": 1,
      "question": "What are the benefits of AI in software engineering?"
    },
    {
      "id": 2,
      "question": "What risks does AI introduce to software engineering jobs?"
    }
  ]
}

---

2. Research

The Research Agent:

• searches the web
• retrieves sources
• extracts relevant evidence

---

3. Claim Extraction

The Claim Extraction Agent converts evidence into structured claims.

Example:

{
  "claim": "AI automation is reducing demand for junior software engineers.",
  "source_url": "...",
  "claim_type": "prediction"
}

---

4. Semantic Storage

Claims are embedded and stored in ChromaDB for:

• retrieval
• similarity search
• contradiction analysis

---

5. Verification + Contradiction Analysis

The system evaluates:

• supporting evidence
• contradicting evidence
• semantic consistency
• stance relationships

using:

• embedding similarity
• NLI transformers
• LLM reasoning

---

6. Final Synthesis

The Writer Agent generates:

• executive summaries
• key findings
• contradiction reports
• confidence explanations

---

🧠 ML Components

Embedding Similarity

Used for:

• semantic retrieval
• claim clustering
• contradiction candidate filtering

Model:

all-MiniLM-L6-v2

---

Stance Classification

Used to determine whether evidence:

• supports
• opposes
• is neutral toward claims

Implemented using:

DeBERTa / MNLI-based transformer models

---

⚡ Performance Optimizations

The system includes:

• Claim deduplication
• Topic clustering
• Batched embeddings
• Async research execution
• Multi-layer caching
• Similarity filtering before reasoning
• Reduced token usage via evidence snippets

---

🚀 Deployment

The backend is deployed using:

• Docker
• Docker Compose
• Nginx reverse proxy
• AWS EC2

CI/CD is automated via GitHub Actions.

Deployment flow:

Push to GitHub
    ↓
GitHub Actions
    ↓
SSH into EC2
    ↓
Docker rebuild + restart
    ↓
Automatic deployment

---

🔐 Environment Variables

Example:

GEMINI_API_KEY=
TAVILY_API_KEY=
CHROMA_DB_PATH=

---

🧪 Running Locally

Backend

cd backend

pip install -r requirements.txt

uvicorn main:app --reload

---

Frontend

cd frontend

npm install

npm run dev

---

🐳 Docker

docker compose up --build

---

🎯 Example Query

Will AI replace software engineers?

The system:

1. breaks the question into sub-questions
2. gathers evidence
3. extracts claims
4. retrieves semantically related claims
5. detects contradictions
6. classifies stance relationships
7. assigns confidence scores
8. generates a structured research report

---

📈 Future Improvements

Potential future features:

• Debate agents
• Persistent long-term memory
• Live collaborative research
• Streaming reasoning traces
• User-specific research profiles
• Autonomous follow-up questioning

---

📄 License

MIT License

---

👨‍💻 Author

Built as an experimental AI research and reasoning platform exploring:

• agentic systems
• retrieval-augmented generation
• structured AI reasoning
• hybrid ML + LLM architectures