From Embeddings to Entities : A Comparative Analysis of RAG Architectures in Academic Domains

Retrieval-Augmented Generation (RAG) systems are transforming how AI models access and utilize external knowledge, specifically in domainspecific applications such as education. Traditional RAG methods typically rely on vector store retrieval, which excels in semantic similarity but struggles with transparency and structured reasoning. This thesis explores an alternative approach, GraphRAG, which uses knowledge graphs to encode explicit relationships between entities from a given passage, potentially offering improved context relevance and explainability. Through a controlled evaluation involving curated datasets across seven academic disciplines and six question types, this thesis compares the performance, retrieval accuracy, and transparency of GraphRAG and vector-based RAG systems. Results show comparable performance across most metrics, with GraphRAG offering notable advantages in source traceability and structured retrieval. Additionally, this study introduces a domain-specific benchmark dataset to assess RAG systems in educational contexts. The findings highlight the value of structured retrieval in enhancing trust and interpretability in AIassisted learning environments and suggest directions for future research on evaluation methodologies and user interface improvements.