Development of a Question Answering System Over Building Codes using Retrieval Augmented Generation

Building codes establish standards for the design, construction, and safety of buildings, ensuring structural integrity, fire protection, and accessibility. However, they are extensive, complex, and frequently updated, making manual querying time-consuming. A promising solution is a Question Answering (QA) system built on Retrieval Augmented Generation (RAG) or its extension, Multi-modal RAG (MRAG). RAG integrates a retriever with a Large Language Model (LLM), while MRAG employs Vision Language Models (VLMs) capable of processing both text and images. This study first evaluated retrieval methods for the National Building Code of Canada (NBCC), comparing pre-trained and fine-tuned LLMs. Results showed Elasticsearch to be the most effective retriever, while fine-tuning LLMs on NBCC data significantly improved domain-specific response generation. Since RAG struggles with tabular data, MRAG was explored as a means to incorporate tabular information. Different input formats, LaTeX and images of tables, were tested, with image-based inputs performing better, though still limited. To address this, Low Rank Adaptation (LoRA) was applied for fine-tuning VLMs on tabular NBCC datasets. Fine-tuned VLMs, particularly Qwen2.5-VL-3B, showed marked improvements, recording a relative 105 percent performance gain. Finally, fine-tuned VLMs were integrated into MRAG and evaluated on a manually prepared NBCC QA dataset that included both text and tables. In addition to these, MRAG framework based on commercial VLMs were also tested on the same dataset, resulting in multiple MRAG frameworks built with either open-source fine-tuned or commercial VLMs. The effectiveness of each end-to-end MRAG framework was then assessed using a combined metric that considered accuracy, ROUGE score, BERT score, and time. Among open-source models, the fine-tuned Qwen2.5 series achieved the best performance, benefiting from their strong reasoning and visual question answering capabilities. Among commercial models, Gemini-2-Flash and Gemini-2.5-Pro delivered the highest overall scores, 0.69 and 0.67 respectively.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.