AI in Automotive Repair:Building a Data Driven Chatbot for Enhanced Vehicle Diagnostics

Sammanfattning

Generative Artificial Intelligence (AI) and Large Language Models (LLMs) present a promising avenue to augment service delivery and customer satisfaction in many sectors, including automotive repair. The traditional diagnostic systems in this sector, supporting the "Triple C" Complaint, Cause, Correction (CCC) of capturing Complaints, identifying Causes and providing Corrections, often suffer from inefficiencies, such as the under-utilization of insights from historic cases, stored in massive databases containing structured and unstructured data. This results in increased costs and extended vehicle downtime due to repetitive or misdiagnosed issues. The primary objective of this research is to enhance the efficiency and accuracy of automotive repair services by developing a chatbot system which can retrieve relevant CCC information from a dataset consisting of technician service and repair entries. The dataset is sourced from workshops from several countries, and includes technical codes and free form text with vehicle and service descriptions. In order to explore and overcome the infrastructure challenges to implement this system within the organizational setup, this thesis aims to develop and analyze two different chatbot systems, both featuring a Retrieval Augmented Generation (RAG) framework to augment Large Language Model (LLM) outputs. The first system being implemented on-premises, integrates the Instructor XL embedding model, Milvus vector database, and Mixtral 8x7B LLM. The second system operates within the Azure cloud environment, employing the text-embedding-ada-002 model for embedding, Azure AI Search for vector retrieval and GPT-3.5 Turbo as the LLM. Both systems are evaluated based on performance, accuracy, scalability, and cost-effectiveness. The on-premises system is better in performance and cost-effectiveness, however, the cloud based system is better in scalability, availability and using searchable metadata. The latter has been implemented by me. The main impact of this research is demonstrated through its contribution to the integration of AI in automotive services, addressing critical aspects such as data privacy, system scalability, and practical implementation of state-of-the-art AI technologies in an industry-specific context. Recommendations for future research include language support, enhanced interactions, improved evaluation, exploration of hybrid architectural frameworks to combine the strengths of both RAG as well as fine-tuning of LLM, and extended integration with real-time vehicle data systems for a comprehensive service experience.