Improving Retrieval for RAG based Question Answering Models on Financial Documents

In recent years, the emergence of Large Language Models (LLMs) represent a critical turning point in Generative AI and its ability to expedite productivity across a variety domains. However, the capabilities of these models, while impressive, are limited in a number of ways that have hindered certain industries from being able to take full advantage of the potential of this technology. A key disadvantage is the tendency for LLMs to hallucinate information and its lack of knowledge in domain specific areas. The knowledge of LLMs are limited by their training data, and without the use of additional techniques, these models have very poor performance of very domain specific tasks. In order to develop a large language model, the first step is the pre-training process where a transformer is trained on a very large corpus of text data. This data is very general and not specific to a certain domain or field, as well as unchanging with time. This is a reason why LLMs like ChatGPT might perform well for general queries but fail on questions on more specific and higher-level topics. Additionally, a model's performance about a certain topic is highly dependent on how often that information appears in the training data, meaning that LLMs struggle with information that does not appear frequently.