Large Language Models (LLMs) have become, in a short time, a useful ‘assistant’ for professionals across many domains. To enhance their usability in specific domains, Retrieval-Augmented Generation (RAG) has emerged as a promising approach. RAG integrates domain specific data into the generation process, enabling more accurate and contextually relevant outputs (question, answer or document). In this paper, we present initial results from an ongoing project developing a RAG-based framework for accounting practices. The goal is to support practitioners by providing contextualized information that aids reasoning, offers guidance, and enables content generation aligned with professional needs. The presented RAG pipeline, focused on context aware question generation, incorporates domain-specific data collection, structure-aware document chunking (including uniform chunking, targeted table extraction, document–table separation, and semantically aware segmentation), dense retrieval using embedding models, and prompt-engineered question generation. We also constructed a ground-truth dataset of accounting-specific question–answer pairs to evaluate question quality using Hit Score and BERTScore metrics. Our results show that document structure and chunking strategies significantly influence the quality of generated questions, especially when handling heterogeneous accounting documents. Semantically guided chunking produced the most promising outcomes. Furthermore, our experiments underscore the importance of specialized processing for tables.