Evaluating Memory Usage and Processing Time in AI-Driven Document Parsing: Extracting Information From Scanned, PDF, And Excel Formats

Abstract

Recent developments in large language models (LLMs) and retrieval-augmented generation (RAG) approaches have changed information management across industries. These breakthroughs have made it possible to design document interaction systems that are intelligent, efficient, and aware of their context. The use of these technologies in corporate document management is the focus of this thesis, which also introduces a new chatbot solution that improves document retrieval and gives users accurate, context-driven help. Using this strong retrieval framework as a foundation, the Retrieval-Augmented Generation (RAG) technique is utilized to merge the optimal embedding model with five different LLMs. To measure how well these models’ function, we look at how well they generate context-aware responses and how well they match up with user expectations. Metrics such as generating time are also evaluated to determine the effectiveness of the system. This thesis shows how combining LLMs, RAG, and advanced embedding methods can change the way corporate documents are managed by making access to knowledge more reliable, scalable, and quick. This paper illustrates the potential for deploying LLM-powered, RAG-driven systems that enable efficient and contextually relevant user interactions across sectors. This promise is highlighted by explaining the system's architecture, methodology, evaluative metrics, and performance benchmarks.