Authors: Utkarsh Singh
Abstract: Despite the widespread adoption of chatbots for customer service, small businesses continue to lack the resources required to develop and customize such solutions effectively. In this work, we propose an approach that enables small businesses to construct tailored customer service chatbots by fine-tuning an open-weight large language model (LLM) using QLoRA [1] on a domain-specific customer service dataset. The proposed method leverages parameter-efficient transfer learning to adapt a pre-trained LLM to the customer service domain, preserving its general natural language understanding and generation capabilities while specializing its behaviour. To extend the chatbot’s functional scope beyond the fine-tuning corpus, we integrate a Retrieval-Augmented Generation (RAG) [4] module that retrieves relevant product information, warranty details, and installation guides from PDF documentation at inference time. This allows the chatbot to ground its responses in accurate, business-specific information rather than relying solely on parametric knowledge. In addition, we log chatbot conversations and apply sentiment analysis to the resulting interaction data, forming a feedback loop that surfaces unsatisfactory exchanges for human follow-up and informs iterative improvement of the system. Taken together, these components constitute a cost-effective pipeline through which small businesses can deploy personalized customer service chatbots without incurring the cost of large-scale data collection or bespoke system engineering.
