Evals without tokens!
by•
Hey all. I'm really interested in how you evaluate your agents in production.
The standard approach is sample, LLM as judge and possibly a golden dataset. That seems to be the standard setup.
But the challenges include: model drift, who is reviewing the llm and it's not live...
I'm looking for who is doing world class evaluation on agents, would love to hear how the community is doing it.
Matt
43 views
Replies
The reviewer question is the one that got us. We had a review model going over another model's drafts and for a while nothing was checking the review model itself.
So we cut down the list of things it is allowed to have an opinion on. Formatting, required fields, whether a stage actually ran, that is all done by a script now. The model only gets the parts that need an actual opinion, and there is much less left for it to drift on.
Ours is a writing pipeline though, not agents in production, so it may not transfer.
Prefactor
@alieksia so are you guys triaging the results of the script to bring an LLM as judge in?
Prefactor
@alieksia love that idea. Would you mind sharing the framework? Would love to have a look at it.
@matt_doughty An LLM-writer holds the rules for how an article should be written, and a separate LLM-reviewer has its own prompt for what to check. The reviewer sends back concrete notes and they go a few rounds until it passes. The deterministic checks sit in front of that, so the reviewer only spends its judgment on what a script cannot decide.
Me and @artk have wondered whether this is worth building into its own tool. Does it seem like something people want?
Prefactor
@artk @alieksia I think that if there is a way to reduce token cost and make, taking action even more valuable. Then there absolutely is value. Full disclosure, I am the founder of Prefactor, and we are focused on evaluating agents. The core of our first class functionality is tokenless evaluation - quality and risk. My broader focus is to create examples where tokenless evaluation can be layered so that you only bring in an LLM when there is a problem, rather than just sampling etc. A) it reduces cost b) it reduces false positives.
Would love you to check it out and see if your idea could work within our product. Always love having passionate design partners for this stuff.