Retrieval-Augmented Generation for Technical Documentation: a Domain-Specific Chatbot for Firmware Manuals

The project presented in this thesis was carried out in collaboration with Brain Technologies, a consulting company specialized in embedded and control systems engineering. The work describes the development of a domain-specific chatbot designed to assist firmware developers during their projects, by providing a tool to help them retrieve information from microcontroller manuals. Firmware documentation is often extensive and heterogeneous, making the search for specific information difficult and time-consuming. The proposed system aims to simplify this process by combining modern Large Language Models (LLMs) with a retrieval pipeline capable of reasoning over complex technical documents. The thesis first introduces the theoretical background on chatbots, Large Language Models, and Retrieval-Augmented Generation (RAG), forming the basis for the practical implementation. The main part of the work focuses on the development of a RAG pipeline that converts technical manuals into a searchable knowledge base. Great importance is given to preprocessing techniques, since relevant information can be found both in textual and tabular form, requiring it to be extracted in a structured way. Subsequently, semantic chunking and vector embedding are applied to enable efficient similarity search. Different large language models were then integrated and compared to assess their stability, cost, and accuracy. The final part of the thesis discusses the evaluation and usability of the system, highlighting both the promising results obtained and the main challenges that emerged during development, in order to identify directions for future refinement and potential deployment in industrial environments.