✨ AI PDF Answer Validation System

🚀 Overview

This project validates AI-generated answers against a financial loan document (PDF).

For every question–answer pair, the system determines whether the answer is:

✅ SUPPORTED — fully matches PDF

⚠️ PARTIALLY_SUPPORTED — some match, some mismatch

❌ NOT_SUPPORTED — no relevant match found

The detection uses semantic embeddings, numeric extraction, and similarity search.

📁 Project Structure

⚙️ Tech Stack

Component	Purpose
Python 3	Core programming language
PyPDF2	PDF text extraction
SentenceTransformers (MiniLM)	Embedding generation
FAISS	Fast vector similarity search
NumPy	Numerical processing
JSON	Input/output formats

📦 Installation

Install required libraries:

pip install PyPDF2 sentence-transformers faiss-cpu numpy

For Windows FAISS:

pip install faiss-cpu-windows

▶️ How to Run the Validator Step 1 — Navigate to src

cd src

Step 2 — Execute the script

python validator.py --pdf ../input-pdfs/axis_loan1.pdf --qa qa_samples.json --out ../validation_results.json

🔍 Argument Meaning

Argument	Meaning
--pdf	Path to source PDF
--qa	JSON file containing questions & answers
--out	Output file where validation results save

📤 Output Format (validation_results.json)

Each entry looks like:

{
  "question": "What is the sanctioned loan amount?",
  "ai_answer": "The sanctioned loan amount is Rs. 15,00,000.",
  "validation_result": "SUPPORTED",
  "confidence_score": 0.82,
  "supporting_text": "[Page X] ... Facility Amount Rupees: 1,500,000 ..."
}

📸 Screenshots Included

Inside /screenshots, the following proof screenshots are available:

🗂 Project folder structure

🖥 Command-line execution of validator.py

These confirm the application works end-to-end as required.

🧠 How the System Works (Simplified)

Extract text from the PDF

Break it into meaningful chunks

Convert chunks → embeddings (MiniLM)

Convert Q&A → embeddings

Compare semantic similarity

Perform numeric extraction & matching

Generate decision label:

• SUPPORTED

• PARTIALLY_SUPPORTED

• NOT_SUPPORTED

🎯 Submission Summary

✔ Complete folder structure
✔ Full PDF → Q&A → Validation pipeline
✔ Final output JSON included
✔ Screenshots provided
✔ Easy-to-run instructions documented