Hamel Husain breaks down the three-layer eval tool stack for AI products: a database for traces, metrics and automated evaluators, and AI-driven analysis. Push-button solutions are unreliable without deep expertise. Teams should use tools as centralized data stores, build custom interfaces, and keep humans validating that metrics actually match real product problems.
When we talk about eval vendors, there's different layers of these eval tools.
At the most basic layer is like a database that help you log your data into it.
It's like interfaces that help you explore that data. Maybe one layer on top of that is sort of metrics tracking where it allows you to log all these metrics that you might have.
Maybe you have some automated evaluators, maybe you're running some experiments on prompts, things like that. Just helps you log it to sort of keep track of the numbers.
Then at like maybe the highest layer is like, hey, we'll just do it all for you. You just push a button and an AI will go and read your traces and tell you what is wrong.
So that third layer, that's not going to work. It can work if you really know what you're doing and you're very advanced.
For all intents and purposes, like, just forget about it.
You know, you need to be in the loop.
And so what you will end up doing is like, okay, by all means, get one of the tools, use it as a database.
You probably want to build your own interfaces on top of the database.
You probably want to annotate the data in your specific way.
You know, your data is unique and your application is unique.
You don't want someone's generic one-size-fits-all thing.
There's no excuse for that nowadays because it's so easy to create like good interfaces like for reviewing data and you can create like things for annotating data.
You're going to suck the data out and you're going to analyze it also yourself. You might want to do that.
So I would say like use a tool, but don't depend completely 100% on it.
You can use it sort of centralizing, like organizing your data and your metrics. Like I wouldn't ever rely on it to—
Yeah. Just tell you what is wrong.
This is like when we talk about the course, we repeat it so many times 'cause people ask us like, what about the tool?
What about can the tool just do it?
If you really wanna know, it's like, okay, if someone does give you off-the-shelf eval, you should take that as like a signal.
Okay, if someone gives you like a helpfulness score, you can score your traces on the helpfulness score and then do your own data analysis to see if that helpfulness score is actually correlated with actually anything wrong.
9 times outta 10, it won't be.
Sometimes you'll find something and you'll create a different score for yourself, but you see you're still in the loop. I would say the value add is the infrastructure part because the shape of the data, these traces, these metric tracking, you probably don't want to build that yourself, honestly, for product. So I would say you're going to build that anyways in the limit, most likely.
