Autonomous Multi-Agent Research & Fact Verification System

An autonomous multi-agent AI system that decomposes research topics into atomic claims, retrieves live web evidence, verifies factual accuracy, detects source contradictions, and synthesizes structured research reports.

---

📌 Project Overview

Verifying complex claims online means cross-checking multiple sources, spotting contradictions, and organizing findings — all manually. Single-prompt LLMs can't do this reliably without real-time web access and structured reasoning.

The Autonomous Multi-Agent Research & Fact Verification System fixes this with a 5-agent LangGraph pipeline backed by FastAPI, powered by live web search (Tavily), vector similarity search (FAISS + SentenceTransformers), and your choice of Gemini, OpenAI, or Claude.

---

✨ Features

• Stateful Multi-Agent Orchestration — 5 specialized agents coordinated in a LangGraph DAG
• Provider-Agnostic LLM Engine — Switch between Gemini, OpenAI, and Claude per request
• Live Web Search & RAG — Real-time evidence retrieval via Tavily Search API
• Local Vector Embeddings — sentence-transformers/all-MiniLM-L6-v2 (384-dim dense vectors)
• Session-Isolated FAISS Vector DB — Fast in-memory semantic similarity search
• Contradiction Detection — Flags numerical mismatches and reporting conflicts across sources
• MongoDB Atlas Persistence — Async job tracking and historical analysis via Motor

---

🛠️ Tech Stack

Component	Technology	Purpose
Framework	FastAPI >=0.110.0	Async REST API with OpenAPI docs
Language	Python 3.10+	Backend runtime
Agent Orchestration	LangGraph >=0.0.30	Stateful multi-agent DAG workflow
LLM Integrations	LangChain (google-genai, openai, anthropic)	Provider-agnostic LLM clients
Web Search	Tavily Python SDK >=0.3.0	Live AI-powered web search
Vector DB	FAISS CPU >=1.8.0	Similarity search vector index
Embeddings	SentenceTransformers	Local text vectorization
Database	Motor >=3.3.2 / PyMongo	Async MongoDB Atlas driver
Validation	Pydantic v2	Request/response schema validation

---

📐 System Architecture

+---------------------------------------------------------------+
|                     FASTAPI BACKEND                           |
|   CORS Middleware · Pydantic Validation · Exception Handlers  |
+---------------------------------------------------------------+
                              |
                              v
+---------------------------------------------------------------+
|                  LANGGRAPH STATEGRAPH WORKFLOW                |
|                                                               |
|   [START]                                                     |
|      |                                                        |
|      v                                                        |
|  1. Claim Extractor  ──>  2. Search Retriever                 |
|     (LLM Factory)            (Tavily + FAISS)                 |
|                                    |                          |
|                                    v                          |
|  4. Contradiction   <──  3. Fact Verifier                     |
|     Detector (LLM)          (LLM Cross-Check)                 |
|         |                                                     |
|         v                                                     |
|  5. Report Generator                                          |
|     (LLM Synthesis)                                           |
|      |                                                        |
|   [END]                                                       |
+---------------------------------------------------------------+
       |              |              |              |
       v              v              v              v
  LLM Factory    Tavily Search   FAISS + Embeddings   MongoDB Atlas
 (Gemini/OpenAI   (Live Search)   (Local RAG)         (Persistence)
  /Claude)

---

📁 Project Structure

multi-agent-fact-verification/
├── backend/
│   ├── app/
│   │   ├── main.py                         # FastAPI entrypoint, CORS, exception handlers
│   │   ├── agents/
│   │   │   ├── claim_extractor.py          # Agent 1: Atomic claim extraction
│   │   │   ├── search_retriever.py         # Agent 2: Tavily search & FAISS indexing
│   │   │   ├── fact_verifier.py            # Agent 3: LLM-based fact verification
│   │   │   ├── contradiction_detector.py   # Agent 4: Cross-source contradiction detection
│   │   │   └── report_generator.py         # Agent 5: Markdown report synthesis
│   │   ├── api/v1/
│   │   │   ├── analyze.py                  # POST /api/v1/analyze endpoint
│   │   │   ├── health.py                   # GET /api/v1/health endpoint
│   │   │   └── router.py                   # API router aggregation
│   │   ├── core/
│   │   │   ├── config.py                   # Pydantic BaseSettings & env config
│   │   │   └── database.py                 # Async MongoDB connection manager
│   │   ├── graph/
│   │   │   ├── builder.py                  # LangGraph StateGraph assembly
│   │   │   └── state.py                    # AgentState TypedDict & domain models
│   │   ├── schemas/
│   │   │   └── analyze.py                  # AnalyzeRequest & AnalyzeResponse schemas
│   │   └── services/
│   │       ├── llm_factory.py              # Dynamic LLM client factory
│   │       ├── tavily_service.py           # Tavily search wrapper
│   │       ├── embedding_service.py        # SentenceTransformers vector service
│   │       ├── faiss_service.py            # FAISS vector DB service
│   │       └── mongo_service.py            # MongoDB async persistence
│   ├── .env.example
│   └── requirements.txt
├── frontend/                               # React + Vite dashboard (separate team)
├── .gitignore
└── README.md

---

🚀 Getting Started

Prerequisites

• Python 3.10, 3.11, or 3.12
• API keys for Google Gemini (or OpenAI/Claude) and Tavily Search

Setup

# 1. Navigate to backend
cd backend

# 2. Create and activate virtual environment
python -m venv .venv

# Windows
.\.venv\Scripts\Activate.ps1
# macOS/Linux
source .venv/bin/activate

# 3. Install dependencies
pip install -r requirements.txt

# 4. Configure environment
cp .env.example .env
# Edit .env with your API keys

Run

python -m uvicorn app.main:app --host 127.0.0.1 --port 8000 --reload

• API Base URL: http://127.0.0.1:8000
• Swagger Docs: http://127.0.0.1:8000/docs

---

🔑 Environment Variables

Variable	Required	Default	Description
GOOGLE_API_KEY	Yes*	""	Google Gemini API key
OPENAI_API_KEY	Optional	""	OpenAI API key
ANTHROPIC_API_KEY	Optional	""	Anthropic Claude API key
TAVILY_API_KEY	Yes	""	Tavily Search API key
MONGODB_URI	Yes	"mongodb://localhost:27017"	MongoDB connection URI
DEFAULT_LLM_PROVIDER	No	"gemini"	Default provider (gemini, openai, claude)
GEMINI_MODEL_NAME	No	"gemini-3.1-flash-lite"	Gemini model string
OPENAI_MODEL_NAME	No	"gpt-4o"	OpenAI model string
CLAUDE_MODEL_NAME	No	"claude-3-5-sonnet-20241022"	Claude model string
MONGODB_DB_NAME	No	"fact_verification_db"	MongoDB database name

*Required only when using Gemini as the provider.

---

📡 API Reference

POST /api/v1/analyze

Request

{
  "query": "Did renewable energy generate more than 30% of global electricity in 2024?",
  "model_provider": "gemini"
}

Parameter	Type	Required	Description
query	string	Yes	Research claim or question (5–1000 chars)
model_provider	string	No	"gemini", "openai", or "claude"

Response (200 OK)

{
  "job_id": "job_d40409a473",
  "status": "completed",
  "query": "Did renewable energy generate more than 30% of global electricity in 2024?",
  "claims": [
    {
      "id": "claim_01",
      "text": "Renewable energy generated over 30% of global electricity in 2024.",
      "verdict": "SUPPORTED",
      "confidence": 0.95,
      "reasoning": "Ember Global Electricity Review 2025 confirms renewables reached 32% in 2024."
    }
  ],
  "sources": [
    {
      "id": "src_01",
      "url": "https://ember-climate.org/insights/research/global-electricity-review-2025/",
      "title": "Global Electricity Review 2025 | Ember",
      "snippet": "Clean electricity generated 40% of global power in 2024..."
    }
  ],
  "contradictions": [],
  "report_markdown": "# Fact Verification Report...",
  "errors": []
}

---

🧠 How the Pipeline Works

[START] → Claim Extractor → Search Retriever → Fact Verifier → Contradiction Detector → Report Generator → [END]

1. Claim Extractor — Breaks the user query into atomic, testable claim statements
2. Search Retriever — Queries Tavily for live web results, embeds snippets into a FAISS index
3. Fact Verifier — Cross-checks each claim against retrieved evidence; assigns SUPPORTED, REFUTED, or INCONCLUSIVE with a confidence score
4. Contradiction Detector — Identifies numerical mismatches and reporting conflicts across sources
5. Report Generator — Synthesizes all findings into a structured Markdown research report