Abstract

Product recommendations present a critical discovery-bottleneck in online second-hand fashion e-commerce due to the large volume of unique items as traditional search and filtering systems struggle to interpret stylistic language. The aim of this thesis was therefore to propose the Style Assistant, a Multimodal Retrieval-Augmented Generation (MM-RAG)-based Conversational Recommender System (CRS) as a potential solution, by delivering technically robust retrieval and recommendation performance while supporting natural language and multi-turn interactions. Following the Design Science Research (DSR) framework, the Style Assistant artifact was designed, developed and evaluated based on the relevance cycle, design cycle and rigor cycle. The evaluation extended established frameworks through a mixed-method framework combining automated LLM-driven RAGAs metrics with an interactive user experience (UX) evaluation based on ISO 9241-11:2018 usability standards in combination with CRS specific quality dimensions proposed by Jannach (2022). The results demonstrated high system effectiveness, where all participants successfully discovered items they liked and would consider purchasing, while latency was identified as a major limitation. The study concludes that the Style Assistant essentially fulfills its aim, demonstrating significant potential to enhance product discovery in second-hand e-commerce. Specifically, it enables users to discover products by expressing themselves naturally with subjective terms suitable for the highly visual domain of fashion. However, limitations were identified and in order to transition the Style Assistant from prototype to a production ready environment further iterations of designing, implementing and evaluating the system are needed. Future research specifically needs to focus on optimizing pipeline latency, hard filtering constraints to fulfill explicit requirements and integrating user personalization.