Large Language Model-Powered Question-Answering Assistant for Citizen Developers
Improving Low-Code Platform Development: Large Language Model powered Question-Answering Assistant for Citizen Developers of a Data Management Low-Code Platform.
Luke, Lavina, 2023
Art der Arbeit Master Thesis
Auftraggebende
Betreuende Dozierende Martin, Andreas, Pande, Charuta
Keywords Large Language models, Question Answering (QA), RAG, Low-code Platform, Citizen developer, LLM-powered assistant
Views: 27 - Downloads: 5
The study explores using Large Language Model-powered QA-assistant to improve Citizen Developers' low-code platform development, addressing uncertainties and expectations surrounding generative AI. The primary objective of this thesis is to answer the main research question of how an LLM-powered QA-Assistant can improve the development process for Citizen developers in a data-centric, low-code platform.
Following the design science research strategy, first the awareness of the problem is established. The awareness of the problem included an exploration of the Low Code Platforms development cycle, the Citizen Developers and their challenges. Furthermore, it explores the landscape of Large Language Models (LLM), their evaluation and different approaches to adapting LLMs for downstream tasks (focusing on Question Answering).
A suitable design for such a Large Langue Model-powered QA assistant was proposed based on the literature review results and requirements from a real-world Low-Code Platform Provider. A working prototype was developed based on the proposed design in Flowise. The main components of the suggested design were based on Retrieval Agumenataion Generation (RAG) to integrate the specific LCP knowledge into the LLM-base.
The evaluation results of the prototype revealed that the Large-Language Model-powered assistant could potentially improve the development process for Citizen developers. Users who solved predefined tasks with the help of the QA-Assistant, solved the tasks faster than the compare group. Additionally, the experience for all three artefact user was stated positive. The ability of the model to answer follow-up questions and give customized solutions based on the inputs in the prompt where noted as a positive effect. Despite the potential, there were some challenges. Those include understanding the usage and how to prompt to receive the right detailed answers and understanding the answers without any previous domain knowledge of the LPC-domain specific terms. Nevertheless, by using the QA-Bot for a short period, the user developed a better understanding, adopted the style of prompting and experimented with different strategies.
The study also made some recommendations for further development and areas where Large Langue-powered solutions might be included in Low Code Platforms.
Studiengang: Business Information Systems (Master)
Vertraulichkeit: öffentlich