That RAG provides a productivity boost I didn't have to tell you. How nice that Nestlé came to bring a wonderful case on this. They initially started with a simple RAG, to prove that it could work. Soon they switched to multi agent RAG. They are building their RAG there with Langchain and on Azure AI search.
Importance of metadata
What we already knew was also confirmed by them. It is super important to determine what metadata you put along with each chunk (this is something we can do nicely in Drupal with the search_api connector in an interface). This is key to the success of the “search”.
Plugging in other tools
What they call Function calling at OpenAI, they call “AI connectors” at Nestlé.
Scalable
They quickly scaled from 100,000 to 200,000 users.
Man In The Middle
This is normally the name of a hacking attack. They also developed a Man in the Middle themselves. A kind of cookie notification before their AI communicates to the outside world:
“Do you really want to share this question, or the information from this conversation, with the Internet?”.
Very cool, as I had seen a similar implementation on questions.ai. My personal prediction: we are going to see this more and more, this will pretty much become the cookie notification of the AI.
Quality control
Nice also to hear that they really need to do quality control, because when switching a model they don't want to have to retest everything.