PDF Answering AI

Local RAG pipeline for PDF question answering

PDF Answering AI

Overview

Developed under the Artificial Intelligence and Electronics Society (AriES), this project implements a complete pipeline for answering user queries from PDF documents without relying on external APIs.

Project Details

Duration: May 2024 – June 2024
Affiliation: Artificial Intelligence and Electronics Society (AriES), IIT Roorkee
Role: Student Project

Objectives

Create a working pipeline to perform the task of answering user queries and questions from a PDF, with minimal use of internet and without external APIs.

Technical Stack

Core Technologies

  • Language Model: Meta AI’s LLaMA3-8B model
  • Optimization: Unsloth library for efficient inference
  • Framework: Langchain for pipeline orchestration
  • Embeddings: Sentence Transformer for semantic understanding
  • Vector Database: FAISS for efficient similarity search
  • PDF Processing: PyMuPDFLoader for document parsing

Architecture

Pipeline Components

  1. Document Processing
    • PDF text extraction using PyMuPDFLoader
    • Text chunking and preprocessing
    • Metadata extraction
  2. Embedding Generation
    • Semantic embeddings via Sentence Transformers
    • Vector representation of document chunks
    • Efficient storage in FAISS index
  3. Retrieval System
    • FAISS-based similarity search
    • Context-aware chunk retrieval
    • Relevance scoring
  4. Answer Generation
    • LLaMA3-8B for natural language generation
    • Context-aware response synthesis
    • Query understanding and interpretation

Key Features

  • Local Execution: Runs entirely on local hardware without external API calls
  • Efficient Inference: Optimized using Unsloth for faster response times
  • Semantic Search: Advanced retrieval using dense embeddings
  • Scalable Architecture: Handles large documents efficiently

Implementation Highlights

  • Minimal internet dependency for maximum privacy
  • Efficient memory utilization for resource-constrained environments
  • Modular design for easy extensibility
  • RAG (Retrieval-Augmented Generation) architecture for accurate answers

Applications

  • Academic research assistance
  • Document analysis and summarization
  • Legal document querying
  • Technical documentation navigation

Impact

This project demonstrates the feasibility of building powerful AI-assisted tools using open-source models and local infrastructure, promoting accessibility and data privacy.

Developed as part of AriES initiative to explore practical applications of Large Language Models in document understanding.