Jev for analytics: turn text columns into gold, in plain SQL | MotherDuck

Edmund Using a language model to label one hundred thousand rows one by one was always a little absurd. This is the first version of that pitch that sounds like an actual analytics feature.

Geffen Yeah.

Geffen Because the claim isn't really 'A I in S Q L.' It's narrower. If the job is a bounded decision over a giant text column, stop generating prose and score a fixed menu instead.

Edmund Right, and that's why this one lands for me. The article's basically saying the gold in the text column is often just a missing typed field, and you're paying for that mistake every time you query it.

Geffen Mm-hm.

Geffen The evidence is pretty concrete, too. They use one hundred thousand C F P B complaint narratives and ask one question: what outcome is the consumer asking for? Then they force it into seven labels like refund, fix credit report, stop contact, identity theft, that kind of thing.

Edmund And the numbers are the whole story here. prompt underscore jev over the full hundred thousand rows in eighty-two point two seconds. But their prompt function with G P T five nano took six minutes nine seconds for only ten thousand rows. That's not a cute margin. That's a workflow change.

Geffen Right.

Geffen And it matches the mechanism. Jev isn't writing an answer token by token. It's doing one-pass scoring over allowed options, so of course it scales differently. This is basically the same thing we were saying with Jev before, just pointed at analytics instead of agent routing.

Edmund Such an Exploring Next disease, by the way. Almost a year of this and we're still getting excited when the answer is, wow, classification was sitting there the whole time.

Geffen You're kidding, but yes. Old systems idea, new model-shaped jacket again.

Edmund Okay, that's good.

Edmund What I also like is they keep the label-design step separate. They sample forty narratives, use MotherDuck's regular prompt function with G P T five mini to propose seven categories, that takes seven point six seconds, then Jev applies that taxonomy to every row. That's a very sane split between expensive thinking once and cheap application many times.

Geffen Sure.

Geffen But that's also exactly where it can break. The author says it pretty plainly: garbage in, garbage out. Jev can only choose from your options, so if your taxonomy is weak, you get a beautifully typed wrong answer. That's my main technical pushback, and it's not even really against Jev. It's against people trusting the neatness of the output.

Edmund Yeah, no, that's fair. The article actually earns some trust by not over-claiming there. He says sampling is a business question more than a technical one, which is right. The hard part is not calling the function. It's deciding whether seven buckets are the right buckets for your org.

Geffen Exactly.

Geffen And the confidence number helps, but only up to a point. A zero point nine five on refund is useful for filtering and group by. It is not proof the taxonomy captured the real intent. This is that old evaluator-under-shift fight again. Clean benchmark, ambiguous production text, same argument, new costume.

Edmund The nice part is the practical use case doesn't need perfection. If I'm an analyst and I can finally do where request dot choice equals refund underscore or underscore charge underscore reversal, or order by urgency score, that's huge. I don't need metaphysical truth. I need a column I can work with before the quarter ends.

Geffen Oh interesting.

Geffen Yes, for triage, segmentation, backlog shaping, even just cutting repeated text parsing costs, I buy it. Where I would slow down is anything adjudicative. If this starts deciding who gets compensation or which complaints are ignored, then no, you need validation on your own messy distribution.

Edmund Which is also why the plain S Q L angle matters more than the model branding. The feature is not 'wow, a proprietary system one model exists.' The feature is an analyst can stay in the warehouse, attach a shared dataset, run one create table as select, and now text became something the rest of the stack understands.

Geffen Mm-hm.

Geffen Also, tiny thing I appreciated: they show the actual attach command for the public MotherDuck share, and they tell you to try limit one thousand first because every call burns A I units. That's the kind of boring honesty that usually means somebody really ran it.

Edmund The boring but works tattoo again. You're never getting out of that bit.

Geffen Nope.

Edmund And honestly, who should care is pretty specific. If you've got giant text columns in a warehouse and the question is repeated and bounded, this is immediately relevant. Customer support, complaints, reviews, maybe messy address-like fields if what you need is decisioning, not a perfect parse. If you're doing open-ended research or inventing a taxonomy from scratch every hour, this is not your tool.

Geffen Right, right.

Geffen I could be wrong, but I think the article mostly holds up because it doesn't pretend Jev replaces language models. It says use the L L M once to design labels, then use the cheaper decision function at scale. That's a much stronger claim than all the vague analytics co-pilot stuff.

Edmund Yeah. This is one of those cases where the practical change is bigger than the conceptual novelty. Not new intelligence, just the right primitive in the right place. Which, annoyingly, is often the whole product.

Geffen Annoyingly for your taste, maybe. For mine, it's perfect.

Edmund Okay, Geffen, let's leave it there before you make me put 'classification in S Q L' on a T-shirt too.