Interactive virtual assistant for addressing motorcycle-related queries using RAG

Abstract

This work presents the design and implementation of an interactive virtual assistant designed to resolve technical inquiries about motorcycles, specifically the Boxer CT100 KS. The system was built using a Retrieval-Augmented Generation (RAG) approach, combined with large language models (LLMs), and operates entirely locally through a web interface with a 2D avatar. The knowledge base was generated from technical manuals, which were processed and stored in a vector database. Multiple combinations of embedding and generative models were evaluated using frameworks such as RAGAS and DeepEval, applying metrics like faithfulness, context precision, and answer relevancy. The results allowed the identification of optimal system configurations, where the best ones excelled in key metrics —such as the sentence-transformers embedding model combined with the Llama-3.3-70b language model, achieving a faithfulness score of 0.964 and context precision of 0.971 in RAGAS-Mistral, and the intfloat/multilingual-e5-base embedding model with Llama-3.3-70b, reaching an answer relevancy of 0.971 in DeepEval-Llama3—, demonstrating the feasibility of customized and private AI-based technical assistance solutions. Improvements are proposed through the incorporation of multimodal capabilities and the expansion of the technical corpus.