jergus

molding search into useful things at

[jergus @jerguslejko]

@gingerhendrixai where to

[jergus @jerguslejko]

@gingerhendrixai where from

[jergus @jerguslejko]

@AgnoAgi what are some coool usecases people built on this?

[jergus @jerguslejko]

@mlejva this resonates! my claude randomly spends minutes bamboofeling, so i don't care about super fast sandbox startup *that* much

[matospiso @matospiso]

You know how the old saying goes…

P.S. you can get these nice shiny stickers at @aiDotEngineer next week, just find @marek_galovic and @jerguslejko at our booth

[Marek Galovic @marek_galovic]

"Just let the agent access our documents."

Sounds simple, until you try to build it. Here's the entire system hiding behind that sentence - and how you can ship it in a day with one tool 🧵

[jergus @jerguslejko]

@AliZYousuf @a_musingcat @tekbog wdym folded?

[will brown @willccbb]

just finetuned gpt-2 to fable-level performance on terminalbench. they both get 0%

[jergus @jerguslejko]

We built a Postgres-compatible SQL layer for TopK.

Our existing query API is already a DataFrame-style API, so SQL fits naturally:

topk Python -> topk SQL
----------------------------------------------------------
select("name") -> SELECT name
filter(field("age").gt(3)) -> WHERE age > 3
fn.semantic_similarity(...) -> semantic_similarity("...")
fn.vector_distance(...) -> vector_distance(...)

SQL is another interface to the same TopK engine, fronted by Pgwire — the Postgres wire protocol. That means any Postgres client "just works":

```
psql "host=https://t.co/bg6AKMTmyO password=<api-key>"
```

I love how simple it turned out! Feels more native than anything else IMHO.

[topk.io @topk_io]

Most search engines force your team to learn a new query language.

Today, we bring hybrid, multi-vector search to the language your team already knows.

SQL becomes your Search Query Language
🧵

[jergus @jerguslejko]

SQL!

[topk.io @topk_io]

Most search engines force your team to learn a new query language.

Today, we bring hybrid, multi-vector search to the language your team already knows.

SQL becomes your Search Query Language
🧵

[topk.io @topk_io]

Most search engines force your team to learn a new query language.

Today, we bring hybrid, multi-vector search to the language your team already knows.

SQL becomes your Search Query Language
🧵

[jergus @jerguslejko]

@RyanMarganti @garybernhardt ah so you mean that pattern when I don't know how to make my code nice so I call it a "factory" righttt

[Jan Curn @jancurn]

Since there's no official @aidotengineer world's fair SF afterparty, @apify, @joinmassive, and @topk_io are joining forces to host an unofficial one on Thursday, 2nd July in South Park https://t.co/ibzLbo01Ft

[jergus @jerguslejko]

@DrizzleORM @slvssb but why

[jergus @jerguslejko]

@calcsam @neondatabase @tursodatabase which one do you recommend for RAG?

[Marek Galovic @marek_galovic]

Single-vector retrieval is cheap, but throws away important detail. Multi-vector keeps it at 10-100x storage and compute cost.

Join me, @HamelHusain, and @isaac_flath to learn how we made multi-vec search practical in production with @topk_io.

Sign up👉 https://t.co/qdG6BNUzhm

[jergus @jerguslejko]

you either die a vector DB, or live long enough to ... ?

[George @georgecurtiss]

All vector dbs are gonna die except turbopuffer.

You’ll still get low latency in memory ones that are opensource, but as a business tpuffer will be the only unicorn.

[George @georgecurtiss]

All vector dbs are gonna die except turbopuffer.

You’ll still get low latency in memory ones that are opensource, but as a business tpuffer will be the only unicorn.