Files
Abstract
Current large language models (LLMs) have demonstrated abilities that, just a few short years ago, would have seemed impossible e.g., question answering. While LLMs like OpenAI’s GPT can do impressive unanticipated things, to maximize their value, the models need to be trained on or have access to additional, often proprietary, data. I compare two popular methods, fine tuning and context injection (a specific application of RAG), for integrating additional data into the LLMs for use in the task of question answering. A suite of semantic measurements is evaluated for use in comparing the answers generated by the methods. I use the best performing measurement, Ada 002 with Cosine Similarity, to show that context injection, using vector embeddings and semantic search, generates answers that are semantically closer to the desired answers, while lacking hallucinations or confabulations. We also provide qualitative and stylistic observations from the experiments further segmenting the two methods.