Integrating Ontologies and Large Language Models to Implement Retrieval Augmented Generation (RAG)

Abstract

. Large Language Models have captured the imagination of the public and the technical community. As powerful as they are they have problems that prohibit their use for highly skilled users. These issues are hallucinations, bias, black-box reasoning, and lack of domain depth. One of the most popular architectures to alleviate these problems is Retrieval Augmented Generation (RAG). In a RAG architecture the LLM is utilized to generate vectors and to parse and generate natural language. The knowledge base for a RAG architecture is typically a set of documents focused on a particular type of vertical (question answering) or horizontal (domain) set of use cases as opposed to the general knowledge base of an LLM. Typically, the corpus for the RAG knowledge base is stored in a relational database. This project investigates the use of an ontology and knowledge graph to form a domain specific knowledge base for RAG in order to leverage LLMs for specific domains without the four problems that typically make them inappropriate for mission and life critical domains. The domain is support of dental clinicians in India who face specific problems that can be significantly improved by better, timely, and easily accessible access to the latest knowledge on dental material products. We demonstrate that using an ontology and knowledge graph to implement RAG has several benefits such as rapid agile development and retrieval by reformulation browsing.