Dalarna University's logo and link to the university's website

This thesis explores the use of retrieval-augmented generation (RAG) for enhancing large language models (LLMs) to answer domain-specific questions about Positive Energy Districts (PEDs). PEDs represent an innovative approach to achieving energy surplus and reducing greenhouse gas emissions, aligning with the European Union’s goals for sustainable urban energy solutions. However, existing Question Answering (QA) frameworks lack the specialized design needed to tackle the complex challenges of PED projects. By constructing a structured corpus of academic literature about PEDs and fine-tuning models such as BERT and T5, the study evaluates their performance answering domain specific questions alongside GPT-enhanced pipelines and ChatGPT. Expert evaluations and cosine similarity analysis reveal that GPT-based pipelines outperform standalone models inaccuracy, relevance, and readability. However, results also highlight limitations in semantic similarity as a sole metric for assessing response quality. This work highlights the role of structured domain-specific corpora and generative models in improving decision-making processes for stakeholders in urban sustainability. The findings suggest future research opportunities, including integrating real-time information retrieval and expanding knowledge bases to refine QA frameworks further. These contributions advance the implementation of energy-positive urban environments and support global efforts toward sustainable urban development.