Institutsseminar/2026-05-22-Room010

Aus SDQ-Institutsseminar
Termin (Alle Termine)
Datum Freitag, 22. Mai 2026
Uhrzeit 14:00 – 14:45 Uhr (Dauer: 45 min)
Ort Room 010 (building 50.34)
Prüfer/in Anne Koziolek
Webkonferenz
Vorheriger Termin Fr 24. April 2026
Nächster Termin Fr 12. Juni 2026

Termin in Kalender importieren: iCal (Download)

Vorträge

Design and Evaluation of a RAG-Based Chatbot for Vitruvius
Vortragende(r) Lorenz Moser
Vortragstyp Masterarbeit
Betreuer(in) Raziyeh Dehghani
Vortragssprache Englisch
Vortragsmodus in Präsenz
Kurzfassung In modelling-heavy software-engineering projects, the relevant knowledge to understand the system is fragmented across heterogeneous source types such as research papers, project documentation, and source code. Vitruvius is one such project: an open-source framework for view-based software development and the technical core of the Convide Collaborative Research Centre. This fragmentation makes the onboarding of new users expensive. Therefore, this thesis develops a Retrieval-Augmented Generation (RAG) based chatbot for the Vitruvius domain in two variants: Naive RAG retrieves purely on semantic similarity, while Graph-Retrieval-Augmented Generation (GraphRAG) additionally exploits structural dependencies in the project’s knowledge base. The contributions are a knowledge organisation pipeline that handles each source type with its own chunking and graph- construction strategy, and the two retrieval variants built on it. To evaluate both variants, this thesis further contributes a 33-question Vitruvius-specific evaluation dataset and a joint evaluation combining a technical sweep, two pilot user studies with Vitruvius users (𝑛 = 5), and an alignment analysis between automated metrics and user perception. The identified Pareto-best GraphRAG configuration improves the mean of all retrieval metrics, such as recall@k and precision@k, by 4–21% over Naive RAG, The user study confirms that the sources retrieved by GraphRAG are more relevant: 3.2 to 2.9 on a 5-point Likert-scale. However, the retrieval-side advantage over Naive RAG comes with a 75% increased end-to- end latency and does not translate into measurably better generated answers. The alignment analysis only shows a weak correlation across all covered metric pairs (𝜌 < 0.20).

Neuen Vortrag erstellen

Bitte melde Dich mit Deinen KIT-Zugangsdaten an. Falls Du trotz Anmeldung diese Meldung siehst, bitte eine/-n Wissensmanager/-in darum, Dich zur richtigen Gruppe hinzuzufügen.

Hinweise