What executives need to know about knowledge management, large language models and generative AI

Abstract

This paper discusses the opportunities and risks presented by large language models (LLMs), which power the popular and widely adopted Chat-GPT types of applications. The potential benefits include support for enhancing the customer journey and efficient management of an ever-increasing volume of information for employees. Risks include hallucinations (made up answers by generative AI that are not factually correct), exposure of corporate intellectual property (IP) to training models, lack of traceability and audit trails and misalignment with brand guidelines. The approach to handling risk described in this paper is retrieval-augmented generation (RAG), which references corporate knowledge and data sources in order to identify precise answers and retrieve exactly what users want. The paper also outlines the need for a knowledge architecture which enables enriched embeddings into vector databases which retain the context of intelligently componentised content. Using RAG requires knowledge hygiene and metadata models, and the paper discusses an experiment in which results were measured with and without the knowledge architecture. The improvement was significant: 53 per cent of questions were answered correctly without the model versus 83 per cent with the model. The use of RAG virtually eliminated hallucinations, secured corporate IP and provided traceability and an audit trail.