Enhancing large language models for data analytics through domain-specific context creation

Abstract

This thesis was written to see the current situation in how the results and outputs of large language models (LLMs) can be enhanced by creating context. The two methods researched were prompting and retrieval augmented generation (RAG). These were chosen because prompting is seen as an easy-to-do solution and RAG is thought to decrease model hallucinations effectively. The thesis also covers how LLMs can be used for data analytics related tasks.

The results show that prompting is a bit more complex than usually thought of. Providing proper prompts was seen to reduce hallucinations, but then again, providing the prompts is not easy and requires planning and possibly multiple trial-and-error runs. Of the two methods RAG proved to be more effective for reducing hallucinations and providing better results. The big downside of RAG was found to be that it can get costly to implement and require experts to do it.

These results showed that currently there are effective methods for enhancing large language models, but further research is required on the topic.