Marcel: A Lightweight and Open-Source Conversational Agent for University Student Support

TL;DR

Marcel addresses the need for fast, grounded, and privacy-preserving admissions information for university applicants by deploying a retrieval-augmented generation (RAG) system grounded in university resources. A key innovation is an admin-steered FAQ retriever that maps user queries to curated FAQs linked to source documents, paired with an efficient on-premise deployment and open-weight generators. The study reports substantial gains in retrieval quality from the FAQ retriever (MRR up to 75% higher than baselines), analyzes the impact of second-stage reranking, and evaluates quantization effects on latency, complemented by real-world deployment insights. The work demonstrates practical, configurable, and scalable practices for resource-constrained universities, with a clear path toward improved information access and reduced staff workload, while transparently discussing limitations and ethical considerations.

Abstract

We present Marcel, a lightweight and open-source conversational agent designed to support prospective students with admission-related inquiries. The system aims to provide fast and personalized responses, while reducing workload of university staff. We employ retrieval-augmented generation to ground answers in university resources and to provide users with verifiable, contextually relevant information. We introduce a Frequently Asked Question (FAQ) retriever that maps user questions to knowledge-base entries, which allows administrators to steer retrieval, and improves over standard dense/hybrid retrieval strategies. The system is engineered for easy deployment in resource-constrained academic settings. We detail the system architecture, provide a technical evaluation of its components, and report insights from a real-world deployment.

Figures (6)

• Figure 1: Overview of student-facing UI. Welcome and consent are shown on first use. For each generated answer, Marcel provides the list of source documents which are linked to the respective websites. Users can rate individual responses using a thumbs up/down button. The overall conversation can be rated after three interactions on a 10-point Likert scale. Marcel abstains from answering questions that are not reflected in the underlying knowledge base. In those cases, no list of source documents is shown. An admin-facing UI is given in \ref{['fig:admin-interface']}.
• Figure 2: Example where a user wants to determine if their academic background meets admission requirements.
• Figure 3: Examples where the chatbot abstains from answering due to knowledge gaps.
• Figure 4: Example of a contextualization issue. The third query is a follow-up on the second query. Omission of the subject ("circumstances when B1 or C1 is required") lead to retrieval/generation of unrelated information.
• Figure 5: Example of a link-seeking question. The generator cannot embed links to sources in the main text (second answer), unless some retrieved context happens to mention the link (third answer).