Energy Advice Made Simple : AI-Powered Recommendations for Property Ownersthrough Retrieval-Augmented Generation

Sammanfattning

This project explores how artificial intelligence, specifically large lan-guage models (LLMs), can be used to help property owners identify waysto improve energy efficiency in their buildings. Our main objectives aretwofold: first, to investigate how prompting can be made more accessi-ble and intuitive for users unfamiliar with LLMs, and second, to comparethe performance of standard LLM prompting with a Retrieval-AugmentedGeneration (RAG) approach. RAG uses a dataset with relevant informa-tion instead of relying on the model’s inconsistent training.In collaboration with Region ¨Orebro County Energy Agency, we devel-oped a web-based solution that allows users to input specific informationabout the building they have in mind using a guided interface and re-ceive relevant tips and information about energy efficiency. The backendemploys either direct LLM prompting or a RAG pipeline that retrievesrelevant context from a curated energy efficiency database before gen-erating a response. The system architecture includes a Node.js server,Google Gemini for language generation and embeddings, as well as anAzure-hosted vector database for semantic search.By comparing the quality, relevance, and usability of outputs fromboth approaches, we aim to assess whether specialized prompting andRAG can provide more accurate and reasonable energy-saving advice thangeneric LLM usage. Evaluations from the energy development departmentat ¨Orebro Region showed minimal performance differences between thetwo methods, likely because energy efficiency is a well-documented topicalready covered in LLM training. However, RAG may offer greater bene-fits in fields with limited available data. Our results demonstrate that AI-generated advice can be made accessible and useful even for non-experts,supporting the goal of promoting informed energy efficiency decisions atscale.