I think I just cooked something 🔥 jev(): a PostgreSQL extension that searches your whole database in natural language. No index, no embeddings, just one function. WHERE jev(people, 'could work from home') or WHERE jev(people, 'name sounds european') 129 rows judged in ~1s for $0.0009. Second run: 6ms from cache.

Livedemo:

this shit is insane.

I know right Will have to calc the query costs on a real, big data prod db though

I hear you, 129 rows isn't much. Updating right now to ~4.500 rows! Will gradually increase to see how far we can take it. Will go live any minute! Oh and please star it on GitHub if you like it 🫶

Yeah I spent about 2 hours brainstorming how I would even do this and here you go building it… great work

Why didn’t you just chat a bit with Claude or gpt?

I did brother. I’m not a dev. I’m a quite retarded

Isn’t that actual extremely expensive? How do you deal with 10m rows?

Super nice, but not always accurate.

Oh my goshhhhh, this is so cool In exploring concepts for our launch demos, I made several where Jev was in the terminal and other similar things, but never thought to put it in an SQL context!

See? you should hire me

Have you tried scaling up from 129 rows? Would be curious how speed and cost scale

Rushed it tbh, will focus on it later and try it out

It’s a cool idea!

Update: Scaled it by 30x. jev didn't even flinch haha

Numbers?

120 rows scan in 1s, is that not slow?

This is brilliant. Hope to see more jev() embedded logic along regular data flows.

That's crazy

typesafe lowkey engineering these sorts of posts as a part of their gtm strategy

logical expression search query -> natural language search query is so powerful. BANGER!

This is nuts

Following , keep us posted brotha

Will do!

129 rows could be a txt file 😂

What if you want to input data to Jev from a joined table?

noted

@grok how does it work with a classifier? Does it pass all the records through jev to say yes or no? One by one or bulk?

for reranking yes, if more than 1000+ row don't think this will work at scale?

It does! Check the live demo at 4500 rows :)

That is very cool.

Real VibeSQL

129 rows 1 second? Isn't it just better to use a vector database?

Actually it's fucked up that I still can't use jev

What’s building the final query?

So now @PlanetScale needs a JIN index

cool 👀👀

What if you give them nothing, no table, anything?

129 rows… sure

Haven't played with it, is it literally just reading the one row, or traversing the data model?

This is brilliant. I’ve been trying to figure out good use cases where it actually makes sense. I think good data plus jev has the opportunity to inform decisions in a really powerful way. I think the one thing I’m still unsure on is how to validate that jev’s opinions are “good”.

Just curious 1. is it FTS + jev? 2. won't it cost more if there is millions of rows in table?

Would you like to give Gliner2.5 a try, sir?

Genius

129 rows lol

What are we witnessing!