Abstract
In medium-sized organizations, frequent turnover of first-level support agents can lead to challenges for new agents, who struggle to discover existing and relevant documentation that would help solve user issues due to inexperience. Consequently, these agents escalate tickets to second-level support professionals, increasing their workload. A proactive knowledge discovery and assistance system targeting first-level service desk agents could help by analyzing tickets using a large language model (LLM) and then finding and presenting relevant documentation utilizing Retrieval-Augmented Generation (RAG) techniques. However, when working with cloud-based LLMs on inference tasks involving sensitive information, data sovereignty is compromised, and there is a risk of confidential content from tickets being leaked, as local information is transmitted to the cloud for processing. To address this issue, we constructed a system based on local LLMs so that the operation of the system does not compromise the privacy and confidentiality of ticket content and wiki documentation, keeping all sensitive data on-premise and secure. Our system, Doku-Assist, proactively finds and presents documentation to first-level support agents, thereby assisting with issue resolution without replacing the human agent. It integrates a DokuWiki-derived knowledge base with the ticket system Znuny. For the evaluation of our system, we used artificial tickets, documentation, and customer issues (to address privacy concerns) based on real-world experience. A second-level support agent was tasked with assessing the utility of the developed user interface as well as the documents proactively discovered and presented, concluding that the found documents presented by the Doku-Assist are useful to proactively fill the knowledge gap of new first-level service desk agents. We conclude that data privacy and law compliance can be achieved by utilizing local LLMs.