Work

Generatiivisen tekoälyn mahdollisuuksien analysointi käyttäjävaatimusten prosessoinnissa

Year: 2024
Type: dissertation
Abstract: The development of various generative AI (genAI) and large language models (LLMs) has been rapid in the 2020s. Although development of these LLMs has its roots dating all the way back in the 1950s only in the latter parts of 2010s started the public’s interest to raise in these modes with the introduction of new processing frameworks. Due to these developments, an ever-increasing amount of people and organisations are looking into finding new ways on how to use LLMs to solve problems in consumer and business context alike. This thesis and empirical research focus on the possibilities on how to implement LLMs as part of an organisation’s customer requirements engineering (RE) workflow. One of the greater problems with the current LLMs considers the recommendations that these tools can have thus this research in this thesis focuses on how to use retrieval augmented generation (RAG) to elicit better prompt answering capabilities without the need for LLM retraining. While implementing LLMs as part of company processes, it is crucial to consider the different opportunities of these systems and the implications of these to the company processes. Since numerous different LLMs available on the market with different operating models ranging from cloud operated closed models to locally operatable open-source ones able to be operated on premises. From a business cost perspective, it is crucial to know how the capabilities of such wide range of different AI models compare to each other in practical applications such as RE. Empirical research conducted in this thesis presents an avenue on how a LLM assisted RE workflow could work in a corporate setting. Moreover, the comparison between a small locally operated model and a cloud operated one was conducted in the field of RE. Findings from this research were encouraging as with RAG implemented with smaller models it was possible to reduce the hallucinations with smaller locally operated models and also get the performance to a similar level to a more capable cloud operated LLM. These findings give a glimpse into the future where it is likely that different companies are using smaller LLMs in specific places in their process workflows as part of the company operations.
Author: Vasama, Aaro
Institution: Aalto University
Language: English
Open Access status: green
OpenAlex ID: W7056494463
Cites: 93
Cited by:
Related to: 10