The integration of large language models (LLMs) in domain-specific applications has been limited due to high computational costs, and the need for expensive and challenging training datasets. This thesis explores Retrieval-Augmented Generation (RAG) and Graph-RAG pipelines to enhance question-answering precision in geothermal energy, addressing these challenges while optimizing computational efficiency. In this thesis, a domain-specific RAG pipeline for geothermal energy is firstly developed by fine-tuning an open-source classifier and embedding model to improve information retrieval. The RAG pipeline uses an open-source LLM to address concerns over proprietary models. The classifier effectively filters relevant geothermal data, increasing domain focus, while the optimized embedding model enhances retrieval accuracy. The results demonstrate that applying RAG improves question-answering accuracy from 55.5% using an untrained embedding model to 72.5% with a fine-tuned embedding model. Additionally, the fine-tuned classifier achieved over 99% precision in classifying text based on context. Meanwhile, the study highlights the environmental impact of increased computational demands, emphasizing the trade-offs between retrieval accuracy and CO2 emissions. A Graph-RAG approach, which enhances RAG by integrating structured relationships between entities, is then employed to improv contextual understanding. Unlike traditional RAG, which relies solely on similarity-based retrieval, Graph-RAG incorporates concept relationships to refine responses. The study evaluates Graph-RAG’s performance in geothermal energy question-answering tasks and demonstrates a 13% improvement in precision compared to RAG, particularly when retrieving fewer nodes and relationships. Moreover, Graph-RAG reduces computational costs by achieving similar accuracy to RAG while using 35% fewer input tokens. This ii improvement comes from Graph-RAG’s ability to leverage nodes and their relationships to better understand the concept. The study further reveals that Graph-RAG is more resilient against misleading statements by cross-referencing nodes and relationships between concepts. This research contributes to the advancement of AI-driven information retrieval in energy engineering by demonstrating the effectiveness of RAG and Graph-RAG pipelines. The findings highlight the benefits of structured entity relationships in improving precision, reducing computational costs, and optimizing knowledge retrieval. The thesis concludes that Graph-RAG offers a more efficient and reliable approach for domain-specific question answering, paving the way for future applications in geothermal energy and beyond.