# FEVER Fact-Checking System with RAG

An end-to-end fact-checking pipeline using Retrieval-Augmented Generation (RAG) to verify claims against a 17.1M sentence Wikipedia corpus. Combines FAISS dense retrieval with a fine-tuned DeBERTa-v3 verifier model.

## Overview

This system verifies factual claims by:
1. Retrieving relevant evidence from Wikipedia using dense vector search (FAISS)
2. Scoring and filtering evidence based on claim-type patterns 
3. Classifying claims as TRUE, FALSE, or NOT ENOUGH INFO using a neural verifier (deBERTa-v3)

Built for the FEVER (Fact Extraction and VERification) dataset with modifications for real-world claim types.

## Architecture
Input Claim - FAISS Retriever (BGE embeddings) - Evidence Quality Scoring (pattern-based filters, using claim type, similarity scores, e.t.c) - DeBERTa-v3 Verifier
↓
Prediction: TRUE/FALSE/NOT ENOUGH INFO

## Features

- Dense Retrieval: FAISS indexing with BGE embeddings for semantic search across 17.1M sentences
- Pattern-Based Filtering: Claim-type detection (authorship, temporal, comparative, geographic) for evidence relevance, along with line similariy scoring and penalization based on sentence text
- FastAPI service with health monitoring and structured request/response
- MLflow integration for hyperparameter evaluation and metrics logging

## Installation

1. Clone the repository
2. Create venv
3. Install dependancies (requirements.txt)
4. Download required data (FEVER dataset, Pre-trained BGE embeddings, Wikipedia corpus; all found from HuggingFace FEVER dataset or downloadable)

Usage:
Training verifier: python src/rag/Verifier/train.py
Building FAISS Index: python src/rag/Retriever/build_index.py
Running inference: 
from test_rag import fact_check
claim = "Tokyo is the capital of Japan"
prediction, response = fact_check(claim)
print(f"{prediction}: {response}")
Starting the API: 
cd api
uvicorn main:app --reload
The API will be available at http://localhost:8000