CEU Electronic Theses and Dissertations, 2025
| Author | Hutter, Jakob |
|---|---|
| Title | Implementing AI for E-commerce Chatbot: From Intent Classification to RAG |
| Summary | This thesis explores the conceptualization and implementation of an AI-powered e-commerce chatbot for a nutrition supplements company, as part of the Sales Data AI project with the University of Applied Sciences of Upper Austria and Heyrise. The goal is to enhance customer satisfaction and drive consumption by providing personalized conversations, product advice, and recommendations. While research exists on chatbot design, intent analysis, sentiment analysis, and Retrieval Augmented Generation (RAG) systems individually, there is a less explored nexus in combining these elements into a cohesive system for specialized domains such as nutritional supplements. Such a specialized chatbot must be conversational, adhere strictly to health information guidelines, and provide precise recommendations. Key challenges include ensuring speed (under 20 seconds response time), accuracy (especially for system-enabling models), and costeffectiveness. A significant challenge is the lack of quality labeled data. Conceptualizing the data flow and implementing necessary machine learning models, this thesis outlines a three-step modular architecture. First, input processing includes intent, sentiment, and language classification, using OpenAI models with function calling for consistency. Intent classification achieved 95.65% accuracy with an average speed of 2.15 seconds. Depending on the detected customer intent, the corresponding data retrieval system is triggered: a standard response, an API call to external sources, or querying a database. Focusing on the non-trivial part of querying a database to create product recommendations for health requests, a vector database for storage has been found to be the optimal solution within the project’s constraints. The data retrieved, along with the intent, language, and sentiment, are then fed into the final part of the architecture of output generation. Mimicking a RAG solution, an LLM (OpenAI’s gpt-4o) was configured with parameters like temperature for consistency and was fed with a structured prompt including the retrieved and classified data as well as company guidelines to answer user requests optimally. A proof-of-concept was successfully deployed as a Telegram bot, demonstrating the robustness of the proposed modular architecture. Performance tests on the primary user intent (Wants Information & Open to Sales Recommendation) confirmed consistent responses under 20 seconds at less than 2 cents per interaction, fulfilling key objectives. This thesis demonstrates the feasibility of specialized chatbot solutions employing functioncalling and one-shot learning, without heavy reliance on labeled datasets. The modular design allows chatbot construction without model retraining and enables seamless updates with emerging AI models, thereby shortening innovation cycles and reducing deployment complexity. These results point to a shift towards more adaptable, cost-efficient, and scalable AI development in specialized e-commerce contexts |
| Supervisor | Battiston, Federico |
| Department | Undergraduate Studies BA |
| Full text | https://www.etd.ceu.edu/2025/hutter_jakob.pdf |
Visit the CEU Library.
© 2007-2025, Central European University