we went to an
@OpenAI
x
@ycombinator
hackathon this week!
@ngalstyan4
and i built an app to iteratively build postgres extensions - it generated a median aggregate function extension successfully :)
I left my job at
@ycombinator
last year to build a startup. Today was our first in-person office hours for W24 with
@gustaf
!
There's so many things that are special about YC, but for me the optimism stands out the most. I'm excited to experience it from the startup side again :)
We just launched
@lanterndb
on
@ycombinator
! 🚀
Lantern is a Postgres vector database that is easy to use, cost-effective, and scales to billions. We support embedding generation, index compression, and external index creation.
The
@ycombinator
retreat ended a few days ago. It was super exciting although I'm just now catching up on sleep.
We're a company built on Postgres, so one highlight was spending an hour talking about Lantern with one of the
@citusdata
founders.
I'm so grateful for this
i don't need a big house.
i don't need a fancy car.
i don't need to travel the world.
i just need the admiration of the two hundred regulars who comment on every hacker news post to support me in whatever i do.
We built external indexing for pgvector in Postgres.
External indexing offloads the resource-intensive index creation process to external machines to reduce the impact on database performance.
Read the blog post from
@D4RK7ET
below 👇
thanks for the shoutout
@TechCrunch
:)
we're building postgres for ai companies. if you're using a specialized vector db or want to improve your postgres performance, please reach out! we'd love to help.
fundraising kicked off about 2 weeks ago, and alumni demo day is tomorrow, so things have been a bit hectic.
excited to be 100% back to building / writing soon, but here's an article from us in the meanwhile :)
Open AI released their new embedding models last month
We worked with
@JoschkaBraun
to evaluate the new models on a healthcare dataset
We used
@lanterndb
for embedding generation / storage and
@PareaAI
for evaluation
Love learning form other YC cos😊
some of
@lanterndb
's engineering updates for april! including
- async tasks
- weighted vector search
-
@ubicloudhq
control plane
- dashboard improvements
Last week we released product quantization (PQ) in Postgres with
@lanterndb
! 🚀
We ran benchmarks over 100M 768d image embeddings. With re-ranking, we got
▶ 90% recall
▶ 73 ms query times
▶ Reduced memory usage from ~300GB => ~60GB
We added client libraries for Knex, Sequelize, MikroORM, and Drizzle for
@lanterndb
! Perform vector search or embedding generation using your favorite JS ORM ☺️
Let us know what you think!
Do you stream vector embeddings into a Postgres instance? This might be relevant to you.
TLDR: a small schema change may make your pipeline 30-50% more efficient. How? - Generate less WAL, use fewer CPU cycles on WAL and applying table / index changes
We support easy embedding generation using
@OpenAI
,
@cohere
, or popular open-source models like
@JinaAI_
.
Generate a text or image embedding inside the database for one-off transactions. For bulk transactions, generate up to 2 million embeddings / hour.
Tell me you’re a systems programmer without telling me you’re a system programmer (my cofounder)
PL I dislike: Makefile
PL I begrudgingly respect: C++
PL I think is overrated: JS/Python
PL I think is underrated: OCaml
PL I like: C
PL I love: Rust
PL I dream of mastering: C++
We built a Pinecone => Postgres export tool!
With a few lines of code, you can export your Pinecone data to Postgres, and build an index using Lantern.
I built a site to browse the Postgres mailing list archives. It supports search using FTS from Postgres + vector search with the Lantern extension.
I've been using it for testing but cleaned it up a bit to share 🙂
This week we worked on docs for using
@lanterndb
with
@llama_index
!
We added an overview of how to get started:
We also added a guide to auto-retrieval:
Would love to learn more about the tutorials people like with LlamaIndex :)
Thank you to our partners,
@gustaf
@bradflora
@dflieb
@garrytan
@daltonc
, as well as others at YC I've gotten to know over the last few years
When I did YC S20, it was fully remote. I'm grateful I got the chance to experience it in-person this time 🧡
The
@GetOnboardAI
team set up a Learn This Repo for Lantern! It's a free tool to ask questions about our source code.
Our source code is available at
Give it a try! ☺️
Introducing , a free tool from Onboard AI.
LTR let's you chat with 300+ of our most popular open source repos for free. From Redis, to Postgres, to NextJS, to LangChain.
Why did we build this?
1. The best way to become a better programmer other than
To use it, just call CREATE INDEX with external=true. Concurrent indexing is supported, and subsequent reindexes also use external indexing
Our pgvector fork to enable external indexing:
Our CLI to run the external indexing server:
customers in 2003: pdfs are bad
customers in 2013: pdfs are bad
customers in 2023: pdfs are bad
founders in 2024: we fixed it!! you can now chat with pdfs… using ai.
customers: what
today's schedule:
8:00am - 2:00pm: try to fix this bug
2:01 - 2:03pm: fix the bug
2:04 - 7:00pm: complain to the people around me about how ridiculous finding that bug was
@eshear
I don’t think this is a proof for that? Both are invalid.
- a plank of wood is not a boat
- adding a plank of wood to something that isn’t a blank of wood doesn’t make it a boat
And yet a boat can be comprised of wood planks.
Something can exist only in the aggregate.
@ShaanVP
I think they're a lens to understand people better - not just the subject but also yourself, your friends, people with similar or dissimilar traits.
Our managed service is currently in beta at – reach out for access!
Or self-host! Our source code is available at
It's been incredible to work with
@ngalstyan4
and
@D4RK7ET
on this. I can’t believe the number of 4am nights 😅
This week we released product quantization (PQ) in Postgres.
For 35M 768 dimension vectors with (m=16), PQ can reduce the index size from ~105GB to ~10GB.
If you’re self-hosting, this can take your monthly cost from ~$1000 to ~$100.
@chefjeffsf
@andrew__rea
I have some conflict about this -- I agree but at the same time for myself these were often fundamentally bad experiences that I don't want to feel "thankful" for
@mralanagon
@ycombinator
@lanterndb
Hey! I'm not aware of Firebase supporting vector search.
AI / Vector support is our core focus. Unlike the others, we built features like PQ quantization, external indexing, inline embedding generation in Postgres, etc. that specifically target AI application needs.
@nooriefyi
@D4RK7ET
We'll be releasing some additional Postgres extensions next week!
For external indexing, we'd love to support other more expensive indexing processes, such as GIN and GIST.