International Journal of Advances in Engineering & Scientific Research
Print ISSN : 2349 –4824
Online ISSN : 2349 –3607
Frequency : Continuous
Current Issue : Volume 12 , Issue 1
2025
Mrs.L.Nivetha, Kowshik G, Thamizhselvan V, Vignesh V
Mrs.L.Nivetha, Kowshik G, Thamizhselvan V, Vignesh V, Department of Computer Science and Engineering, Kongunadu College of Engineering and Technology, Tamilnadu, India
DOI : https://doi.org/10.5281/zenodo.14922126 Page No : 34-44
Published Online : 2025-02-25
Download Full Article : PDF Check for Updates
The exponential growth of unstructured data in PDF format presents challenges for efficient extraction and comprehension. This paper introduces a Retrieval-Augmented Generator (RAG) driven PDF Data Extraction framework leveraging Large Language Models (LLMs) to enhance accuracy and contextual understanding. Our approach integrates retrieval-based augmentation with generative AI, ensuring precise and contextually aware extraction from complex documents. The framework consists of three key components: document preprocessing, retrieval-augmented generation, and post-processing refinement. First, PDF documents are parsed using Optical Character Recognition (OCR) for scanned content and structured parsing for digital text. A vector database (e.g., FAISS, Pinecone) indexes document segments using embedding models (e.g., Open AI’s Ada, SBERT). During query execution, relevant document chunks are retrieved using semantic search and passed as augmented context to an LLM (e.g., GPT-4, LLaMA, Mistral) for accurate extraction and summarization. The output undergoes post-processing using Named Entity Recognition (NER) and rule-based validation to refine extracted information. This hybrid approach enhances information retrieval efficiency, mitigates hallucinations, and enables domain-specific adaptation. Our method outperforms conventional PDF parsers in handling complex layouts, multilingual content, and domain-specific terminologies, making it ideal for legal, financial, and scientific document processing. Experimental results demonstrate significant improvements in precision, recall, and response coherence over baseline methods.
Keywords: RAG, LLM, PDF Data Extraction, Semantic Search, Vector Embeddings and NLP
