<![CDATA[This study presents the design and development of a Retrieval-Augmented Generation (RAG)-based Virtual Boiler Assistant (VBA) integrated with a Large Language Model (LLM) to support boiler-system learning for cadets at Politeknik Pelayaran Surabaya. The system uses FastAPI as the backend framework, ChromaDB as a local vector database, Gemini 2.5 Flash as the generative LLM, and React as the frontend interface. A total of 192 text chunks were extracted from four official training documents used in the steam plant laboratory and Computer-Based Training (CBT) simulator. For each query, the RAG pipeline retrieves the four most semantically relevant chunks using cosine similarity over Gemini embeddings, then injects that context into the LLM prompt to produce accurate responses with source citations. The system was deployed on a VPS with Server-Sent Events (SSE) streaming, achieving an average response time of 15 seconds and an average semantic-similarity score of 0.73. Accuracy was measured using 20 validated ground-truth questions covering four domains: startup procedures, fuel systems, safety mechanisms, and monitoring software. Each response was assessed with a three-point rubric: 0 for incorrect answers, 1 for partially correct answers, and 2 for fully correct answers, all compared against reference answers derived directly from institutional teaching documents. The system achieved 30 of 40 points, or 75%, exceeding the minimum threshold of 70% set at the start of the study. These results indicate that a RAG-based virtual assistant, at least at this testing scale, can serve as a practical supplementary learning tool for cadets in maritime technical training.]]>
