Generative AI presents an incredible opportunity for organizations to gain a competitive edge. So what keeps organizations form realizing its full potential. Maxime Vermeir, Senior Director of AI Strategy at intelligent automation company ABBYY, shares insights into ways to achieve real results using generative AI.
Generative AI presents an incredible opportunity for organizations to gain a competitive edge. However, the sheer volume of data used to train models like ChatGPT or GPT-4 by OpenAI raises some important considerations. GPT-4, for instance, has been trained on an astonishing equivalent of 1.3 million books, a vast amount of information beyond the comprehension of any individual. Yet, what has surprised many is the occurrence of what is known as “hallucinations” in these models. These are instances where the output or answers provided are entirely fabricated, lacking any factual basis, yet appearing remarkably real. There have been documented cases of academic papers being cited, even though they do not exist, and references to US laws being used in arguments without any trace of their existence.
To illustrate this phenomenon, let’s imagine having a conversation with a random person on the street. In this hypothetical scenario, regardless of the question asked, they feel obliged to provide a well-founded answer. However, how could we ascertain their expertise in the specific domain we are inquiring about? Since they didn’t refuse, they would likely create a response based on the knowledge and data they possess, striving to make it sound as credible as possible, as that was the task at hand.
In real-life scenarios, we would never rely on such an approach. Instead, we would seek a domain expert, provide them with contextual information about the question, and even after receiving an answer, evaluate its credibility using rational thinking.
Leveraging generative AI follows a similar principle. Without the right context and specific domain knowledge to guide the technology, it will draw from its entire data set. This is why ‘prompt engineering’ quickly emerged as a practice, where the scope within which the model should operate is specified and provided. We have all encountered cheat sheets stating something like, “I am an XYZ expert…” However, simply providing instructions of this kind falls short when attempting to utilize generative AI within a business setting. The technology requires the insights and knowledge unique to your business to truly deliver the desired answers.
This is where the importance of an extensive business knowledge base comes into play. The knowledge base serves as a crucial component for leveraging a technique called context injection.
To build such a knowledge base, it becomes necessary to leverage other AI tools that can provide the required data. One such technology is intelligent document processing (IDP). Given that approximately 90% of business processes involve documents, leveraging AI models specifically designed to convert images of documents into readable text and extract valuable information becomes imperative. Once this structured data is obtained, it can be used to create an organizational knowledge base that combines the original prompt with the necessary context for the language model to understand the scope, domain, and relevant data. This allows the model to provide knowledgeable answers that are factual, accurate, and, most importantly, useful for the task at hand.