Evaluation for AI products means measuring output quality with 5 to 10 metrics, not one number. Without clear metrics, teams play whack-a-mole, fixing problems after they appear. Systematic evaluation builds trust, reduces churn, and enables continuous improvement.
What is evaluation and why is it very important for LLMs? And here's what we came up with. It is systematic and it is a measurement. So there's some number or set of numbers involved in there, and it measures quality in some way. And this is where it's subjective. Quality means different things to different people. It can mean correctness, it can mean personalization. So these things are what you are going to figure out how to apply for your application and product. Many times it goes beyond a single number. In practice, in most applications, we see people actually having 5 to 10 different evaluation metrics for their products. And this is actually not that crazy to think about, because in traditional software, traditional applications, there are lots of product metrics that are very useful and impactful for the business. And why should you bother with evaluation? For a number of reasons. One is you need your users to trust your product. That's the bare minimum. You want them to come back. You don't want them to churn. So they need— your AI pipeline needs to be reliable, needs to be safe. Safe is— generally interpret that it doesn't mean safety nuclear threats or whatever. It means something literally that a human can trust for your application. And it needs to be useful. So for example, Rust Rebot cannot suggest things that are gonna turn people off. The other thing that's very nice about evaluation is when you have criteria or metrics that you know you want to measure, you can now try to monitor them, hopefully automatically, and then automatically alert yourself when the suggestion or the quality degrades. Ultimately, if you don't have this, you cannot do systematic improvement. Otherwise, it becomes this whole game of whack-a-mole.
