LLM-integrated architecture for document intelligence and semantic retrieval

Abstract

This study presents an Artificial Intelligence-enabled digital transformation system for intelligent enterprise document management. The proposed platform combines Optical Character Recognition, semantic embeddings, vector search, and Retrieval-Augmented Generation to improve document retrieval and knowledge access in organizational environments. Unlike traditional keyword-based systems, it supports context-aware search and evidence-grounded question answering over internal documents. The system is implemented with a practical full-stack architecture using React, FastAPI, and PostgreSQL with pgvector. To improve retrieval quality, we introduce a K-Means-based filtering step that separates highly relevant results from noisy ones before answer generation. Experimental observations show that this hybrid approach provides more focused outputs and better practical precision while preserving strong ranking quality. The results demonstrate strong potential for scalable enterprise knowledge intelligence applications.