Access Type

Open Access Dissertation

Date of Award

January 2025

Degree Type

Dissertation

Degree Name

Ph.D.

Department

Computer Science

First Advisor

Sorin Drăghici

Abstract

The rise of Large Language Models (LLMs) has transformed artificial intelligence, offering advanced capabilities in text generation, natural language understanding, and multi-modal interactions. However, their use as standalone tools or as perceived repositories of static knowledge has limited their potential in real-world applications, especially in critical domains like healthcare and scientific research, where transparency, explainability, and accountability are paramount. This research addresses these limitations by conceptualizing LLMs as reasoning engines within a hybrid framework that integrates retrieval-augmented generation (RAG) and case-based reasoning (CBR) within a note-taking application.

The study introduces a novel system, LmRaC, designed to enhance the reliability, explainability, and collaborative utility of LLMs. LmRaC is embedded within the Obsidian note-taking application, offering a dynamic environment where users can interact with LLM-powered tools to generate traceable and goal-oriented solutions. By combining authoritative RAG-based information retrieval with a collaborative CBR framework, the system ensures that answers are grounded in verified data while fostering iterative problem-solving processes. This design mitigates hallucinations, improves reasoning accuracy, and supports the seamless integration of new quantitative data into evolving experimental contexts.

Across its three development phases, LmRaC showcases innovations that redefine the integration of LLMs into scientific workflows. In the command-line phase, a two-pass RAG approach incorporates a "usefulness" metric to filter retrieved information, reducing hallucinations and enhancing precision. The web application phase introduce a fully interactive user interface to domain knowledge and experimental data and enhanced reasoning to enabling comprehensive, data-aware answers. Finally, the collaborative note-taking phase embeds case-based reasoning (CBR) into a unified platform, leveraging graph databases and vector embeddings for efficient knowledge management while promoting human-AI collaboration through iterative workflows. Together, these advancements create a dynamic, transparent, and adaptive system that amplifies human reasoning and facilitates collaborative scientific discovery.

The system’s interdisciplinary design integrates human workflow principles with LLM-enhanced tools to enhance interaction and adaptability, providing a powerful tool for addressing complex, evolving problems in scientific research. By positioning LLMs as collaborative partners rather than autonomous oracles, this work lays the foundation for a new paradigm in AI-human collaboration, emphasizing transparency, ethical accountability, and dynamic knowledge integration. These contributions set the stage for more responsible and impactful applications of AI across critical domains, advancing human knowledge and decision-making.

Download

Share

COinS