Developing a maritime-specific LLM utilizing RAG architectures
Ανάπτυξη εξειδικευμένου μεγάλου γλωσσικού μοντέλου για τη ναυτιλία με αξιοποίηση αρχιτεκτονικής RAG
View/
Keywords
Retrieval-Augmented Generation (RAG) ; LLM ; Large language models ; Maritime ; Semantic search ; Vector databases ; Information retrievalAbstract
Maritime technical manuals contain extensive operational, maintenance, and safety-related information, making the retrieval of specific technical details a time-consuming task. This thesis examines the use of Retrieval-Augmented Generation (RAG) as a method for improving access to information contained in maritime technical documentation. A locally deployed RAG system was designed and implemented for question answering over maritime technical manuals. The system processes PDF documents into overlapping text chunks, converts them into vector representations using the all-MiniLM-L6-v2 SentenceTransformer model, and stores them in a persistent ChromaDB vector database. For each user question, relevant document chunks are retrieved and used as context for the Qwen2.5-3B-Instruct language model. Both retrieval and response generation are performed locally, without relying on an external language-model API, while a graphical user interface was developed to support direct interaction with the system. The system was evaluated using 24 questions based on four maritime technical manuals. Relevant information was successfully retrieved for 20 of the 24 questions, corresponding to an overall retrieval success rate of 83.3%. Of the generated responses, 15 (62.5%) were classified as Fully Correct, six (25.0%) as Partially Correct, and three (12.5%) as Incorrect. The results show that the proposed lightweight local RAG system can support access to technical information across different maritime manuals, while also highlighting that successful retrieval does not always guarantee a complete or accurate final answer. The system therefore provides a useful basis for further development of locally deployed RAG tools for maritime technical documentation.