By Ferdinand Netsch, October 2024
How to build an AI that can answer all questions about the company, products, and instructions for employees and customers?
Maybe you’ve heard of RAGs or vector databases . These are concepts that use generative AI to search and deliver large amounts of data.
This brings up the question, why not just let the model learn this information (with fine-tuning)?
The crucial difference is that using RAG, the AI model can tell exactly where it got the necessary information from. This is not the case with previously learned information, which is why you can never be sure if it is correct or just a hallucination of the model.
As the technical director of LUNOS Lüftungstechnik GmbH for indoor air systems, Michael Merscher knows best why it is so important to make this information available to the model as easily as possible. Why him in particular?
Because the fans the company offers are complex systems that can raise completely unforeseen questions. For example, one version of the system talks about a switch that has the same function as another version, but with a different label and color. Due to the similarity in the description and the search strategy underlying most AIs, the model cannot determine that the customer’s question is about the same switch and may interpret it incorrectly.
“It takes a lot of hard work. You really have to rack your brain to cover as many potential questions as possible.” Over the past few months, his team has built a knowledge base of more than 2,000 pages.
The knowledge base tries to describe the problem as semantically similar as possible to ensure that the content can be found by the algorithm searching the vector database. “That’s why it doesn’t help to just throw technical drawings, process descriptions, and all sorts of documents into the database. The information really has to match the context, and it has to be documented from what point in time the information is, where it comes from, and what exactly is being described by that information.”
What else is missing? Exactly, a precise instruction! 
“The AI is like a very smart child. I’ve often smiled at what I did wrong when asking questions. The result was correct, but useless for the purpose”.
