In previous studies, large language models (LLMs) were fine-tuned using Korean utterance data to improve performance in small-scale computing environments by applying the low-rank adaptation (LoRA) method. In addition, Gradient Checkpointing and Gradient Accumulation techniques were employed to address computational resource limitations, enabling efficient fine-tuning under constrained computing conditions. The purpose of this study is to develop a marine geospatial information LLM. To achieve this, the LLM previously fine-tuned on Korean utterance data was enhanced by integrating a retrieval-augmented generation (RAG) framework, a document-based inference approach, with marine geospatial information data. First, domain-specific terminology learning was conducted using the International Hydrographic Organization Dictionary (S-32), which provides standardized definitions. Additionally, data on current speed, current direction, wind speed, and wind direction were collected from the Badanuri marine information service provided by the Korea Hydrographic and Oceanographic Agency and incorporated into the RAG knowledge base. Subsequently, S-101 and S-102 datasets were preprocessed to extract bathymetric depth information and were also integrated into the RAG framework. In conclusion, this study demonstrates the feasibility of developing a marine geospatial information-specialized LLM in a resource-constrained environment and enhances the practical applicability of the proposed marine geospatial information LLM through RAG-based knowledge integration.