Ανάπτυξη υβριδικού συστήματος ερωταποκρίσεων με αξιοποίηση γλωσσικών μοντέλων και γράφου γνώσης
Design and implementation of a hybrid conversational AI system using LLMs and knowledge graphs
View/
Keywords
RAG ; NLP ; Γλωσσικά μοντέλα ; Γράφος γνώσης ; Hybrid Q-AAbstract
The rapid advancement of Large Language Models has enabled the development of question answering systems with high linguistic fluency. However, such models exhibit limitations in knowledge-intensive applications, particularly when accurate and verifiable information is required. Autonomous language models may generate answers that are not grounded in factual data, a phenomenon commonly referred to as hallucination.
This thesis proposes and implements a hybrid question answering system that combines Retrieval-Augmented Generation and Knowledge Graphs to improve the accuracy and reliability of generated answers. The system retrieves unstructured knowledge from a document collection using vector-based representations and retrieves structured facts from a Knowledge Graph through Cypher queries. Information from both sources is integrated in a dedicated synthesis stage to produce the final response.
The thesis presents the theoretical background of language models, Retrieval-Augmented Generation, and Knowledge Graphs, followed by the design and implementation of the pro-posed hybrid system. Finally, the limitations of the current approach are discussed, and directions for future improvements and extensions are outlined.