Hint: It’s Critical
We have been accused in the past of hyperbole on how important data is to enterprises looking to transition to digital methodologies. Investing in data is a cultural shift that not everyone will be comfortable with, let alone buy in to as a philosophy. With the advancement in generative AI platforms in 2023, data readiness has taken on a heightened level of importance.
ChatGPT And Retrieval Augmented Generation
Implementing chatbots, personalized search, and other generative AI functions are the standard use cases for generative AI, and the platform most people think about first is ChatGPT. OpenAI created a Large Language Model (LLM) on top of a vector database that relies on existing parsed documentation in order to process results. The LLM allows for processing vast amounts of data very quickly and passing these results in a human-readable format.
But what happens when ChatGPT doesn’t have the data to answer a question from a user? And what exactly happens with the data you push to ChatGPT to feed the prompt? The answer to the first question is simple: AI hallucinates when it doesn’t have the answer. The second question is quite more difficult to answer.
Hallucinations in AI like ChatGPT are well documented. There is the case of the lawyer who asked for court precedents about seatbelt cases, and instead was fed cases entirely made up by the LLM. Google lost $100 Billion in value in one day after it’s AI engine, Bard, responded incorrectly a question about the James Webb Telescope during a demo. A professor in California was part of a potentially liable hallucination when ChatGPT claimed he sexually harassed a student on a class trip to Alaska, using a Washington Post article as proof. The class trip, the sexual harassment, and the Washington Post article were all fake.
Generative AI hallucinates when it doesn’t have enough data to answer a question, when the data that has been used in training is invalid, or when the creativity settings are set too high and allow the AI engine to invent facts. For most businesses, the product data they possess is not part of the LLM’s and therefore cannot be used to generate responses on products it knows nothing about. This is where Retrieval Augmented Generation, or RAG, comes in.