Publication Date

Fall 2025

Degree Type

Master's Project

Degree Name

Master of Science in Computer Science (MSCS)

Department

Computer Science

First Advisor

Ching-Seh (Mike) Wu

Second Advisor

Navrati Saxena

Third Advisor

Robert Chun

Keywords

Teaching assistant, Large Language Models, Retrieval Augmented Generation, Prompt Engineering, Generative A.I., Gemini, Gemma, Llama, GPT-OSS

Abstract

Retrieval Augmented Generation (RAG) in integration with Large Language Models (LLMs) has demonstrated significant improvements in producing factually grounded and context-aware responses. This is one of the most required aspects when developing tools for programming education. As the demand for programming skills has been growing, both students and educators are face challenges in obtaining personalized and reliable assistance. Existing LLM-based solutions are often suffering from hallucinations, poor context retention, and reliance on static knowledge. To address these issues, this study introduces a hybrid RAG-based framework with pre-trained LLMs to enhance contextual understanding and response accuracy. A multi-provider architecture is designed to support Gemini, Gemma, Llama, and GPT model, enabling educators to flexibly select models suited to their instructional needs. The system includes two configurations: one offering students an interactive instructor-like guidance and another assisting educators in automating tasks such as grading, and assignment generation. RAG is supported by the Milvus vector database and optimized through semantic chunking and query decomposition. The proposed framework is tested using prepared CanItEdit [12] dataset. The system achieves significant performance by 20% reduction in hallucination rates and 80% improvement in BLEU scores. These results highlight the potential of context-driven retrieval in bridging the gap between generic AI tools and personalized, adaptive educational support.

Download

Available for download on Saturday, December 19, 2026

Share

COinS