Passionate about teaching | Addicted to learning | AI Lead @supabase 🎥 rabbithole.now 🐰🕳🩺
- Tweets 406
- Following 415
- Followers 3,190
- Likes 3,471
Pinned Tweet
Introducing supabase.sh: Supabase docs over SSH
Give your agents direct bash access to @supabase docs so that they can explore them the same way they do with code
6
25
145
16,521
Excited to introduce postgres.new, the in-browser Postgres sandbox with AI assistance 🚀
34
207
1,458
192,386
For the @supabase folks, introducing the SQL-to-REST API translator 🚀
Translate any SQL query to the equivalent PostgREST request or Supabase client code.
12
15
164
15,920
We just rolled out an exciting new feature on database.build (formerly postgres.new): Live Share
Connect external postgres clients directly to your in-browser PGlite databases
3
17
145
11,049
Excited to share some news: I've join @supabase's engineering team to help lead their #Developer #Documentation site ⚡️ 💻
Sometimes the best opportunities aren’t planned. 🧵
16
5
90
9,310
Matryoshka embeddings allow you to "shorten" their dimensions (eg. OpenAI's v3 embeddings). But if you're like me, you probably want to know:
‣ how does shortening actually work?
‣ how are these models trained differently?
‣ can we take advantage of this in vector search?
4
18
54
17,173
An absolute pleasure working with @samwillis on postgres.new - this is a guy who will obsess over DX until the very end. Give pglite.dev a try!
1k GitHub stars since yesterday 🤯
I could not have had more fun than working with @ggrdson from @supabase on the pglite.dev + postgres.new collaboration 🚀
2
3
45
2,489
There are lots of great guides on AI concepts (vector dbs, embeddings, RAG) - but how do you turn those POC's into a production-ready app?
- What are best practices?
- How do you implement permission-based RAG?
- How do you index embeddings for scale?
Hope this helps!
The missing pieces to your AI app 🧩
A video guide to go from pgvector to prod using Supabase: piped.video/ibzlEQmgPPY
2
5
32
9,191
Series C complete @supabase. It’s been super cool to be a part of this team and watching from the inside how all the pieces come together 🚀 Honoured to be working with such talented folks!
Supabase, a Postgres-centric developer platform, raises $80M Series C tcrn.ch/3BleQwr
33
1,220
Want to try a 100% local LLM without jumping through hoops? 🏃🛟🛟🛟
Meet WebLLM 🤝
github.com/mlc-ai/web-llm
This runs a LLaMa variant 100% local in your browser using WebGPU.
2
2
31
4,876
Database branching is something I'm particularly stoked about. I've always wished there was a turn-key @vercel-like preview that spun up a new environment every GitHub PR - but for the _backend_
Now it's possible 🙌 and both integrated together at the same time (full-stack) 🚀
LW🆇 Day 3 Main Stage announcement
Supabase Branching
⬥ Vercel Previews
⬥ Data Seeding
⬥ Preview branch
and more: supabase.com/blog/supabase-b…
ALT Supabase Branching.
1
3
27
1,485
In case you missed it, @langchain added a new tool called self-querying retrievers where the LLM generates the DB query itself based on your prompt then executes it for you. Think text-to-SQL but safer + automatic execution.
@supabase now supported: python.langchain.com/docs/mo…
The new self-query chain from @langchain is awesome
It takes a user query like “I want to watch a movie rated higher than 8.5” and then can detect that it’s a simple structured query like:
select *
from movies
where rating > 8.5
This will give an incredible performance improvement and more accurate results.
We just added @supabase support:
python.langchain.com/docs/mo…
1
5
23
10,814
Until now, I hadn't given open source LLMs any serious attention - because realistically they just couldn't compete with OpenAI...
...not so true anymore - if you haven’t gone down the OSS LLM rabbit hole yet, I’d say now is the time 🐰🕳️
2
1
26
2,141
Powered by PGlite, the WASM-built Postgres by the brilliant folks over at @ElectricSQL (special shout to @samwillis)
2
1
26
2,569
@supabase Launch Week 12 almost here 🦾
The swag just keeps getting better 🤩
supabase.com/launch-week/tic…
1
2
25
268
Or the blog post, which goes into the nitty-grittys:
supabase.com/blog/postgres-n…
3
1
23
1,892
#SupaLaunchWeek 🆇 is coming up December 11-15 🚀
Grab a ticket for a chance to win @supabase swag:
supabase.com/launch-week/tic…
1
23
2,791
We recently put up new pgvector guides on different types of search (semantic search, keyword search, hybrid search).
supabase.com/docs/guides/ai/…
lmk if you hit any questions/issues, trying to make the vector docs as intuitive as possible.
1
21
599
How @supabase's SQL-to-REST tool works under the hood
(if you have rabbit hole syndrome like me) 🐰🕳️🩺
We just shipped a new SQL -> PostgREST translation tool
Check it out in the @supabase docs:
supabase.com/docs/guides/api…
🚢 @ggrdson
1
1
21
1,263
I’m a big believer in demystifying things, including how day-to-day life works at a remote software company.
Thankfully @supabase is open source, so it’s easy to share these kinds of things 🚀
1
2
21
1,551
Many of us use OpenAI’s embedding models for our projects - but is there any competition? Ideally performs well, easy to run, open source?
(spoiler: very well, easy, open)
Did a deep dive to help answer these questions 🎥
piped.video/QdDoFfkVkcw
2
3
19
3,960
Didn’t make it to the AI Engineer summit? Watch the live stream now.
(pgvector talk by @supabase ceo @kiwicopple in ~1 hour)
5
19
2,091
Ever wished AI had first-class support in your SQL editor? Generate/edit/debug queries for you? Wish no more 😃
Day 3 announcement:
Supabase Studio v3.0:
⬥ AI SQL editor
⬥ Schema diagrams
⬥ Wrappers UI for S3, BigQuery, @ClickHouseDB, and @logflare_logs
Blog post: supabase.com/blog/supabase-s…
2
19
1,765
For a deeper dive, check out the video demo:
piped.video/ooWaPVvljlU
1
18
1,950
Congrats on the launch! Built with @supabase 🚀
I'm super excited to launch ⌘ Langbase.com 🥳
⌘ Langbase – Composable AI developer platform to ship AI features in minutes, not months.
Deploy AI Pipes:
Hook any LLM to any data, hyper-personalized API
AI Memory:
Managed search engine API with RAG tools
1
3
18
1,711
Never thought I'd help revive Clippy (I'm a dev, not a doctor)! Had a blast building this with the Supabase team.
pgvector in action!
Meet Supabase Clippy, ChatGPT for Supabase Docs 📎
Built by @ggrdson.
supabase.com/blog/chatgpt-su…
ALT Supabase Clippy: ChatGPT for Supabase Docs
1
1
17
4,780
Things move fast at @supabase! Already another #SupaLaunchWeek coming up August 7-11 🚀
supabase.com/launch-week/tic…
2
15
3,176
Do you store your OpenAI embeddings in Postgres using pgvector? Any scaling challenges? Switch to an embedding model with fewer dimensions to significantly speed up queries and reduce memory load.
Technical exploration of leveraging embeddings with reduced dimensions to boost performance in #pgvector.
By @ggrdson, @egor_test, and @olirice_dev.
supabase.com/blog/fewer-dime…
ALT pgvector: fewer dimensions are better.
1
1
15
1,803
In case you missed it, you can run aggregate queries (sum, count, etc) directly from @supabase's client libs & API (no RPC wrapping). More details:
supabase.com/blog/postgrest-…
1
1
14
620
#SupaLaunchWeek #7 is coming up where @supabase launches something new every day for a week 🚀📆
In case you like swag, there's a mechanical keyboard giveaway ⌨️
supabase.com/launch-week/tic…
3
13
3,976
Will do my best to capture it all in the YT video!
(thanks for the kind words and attending 💚)
1
12
110
Nomic embed v1.5 is the first open model I’ve seen support Matryoshka embeddings.
Also includes ONNX weights, so works with @xenovacom transformers.js
Nomic Embed v1.5 is on @huggingface and compatible with SentenceTransformers:
huggingface.co/nomic-ai/nomi…
2
1
10
936
As always, all code is open source:
github.com/supabase-communit…
1
10
519
You’re building search for your product 🔍
When should you use sentence embeddings vs full text search? I like to think of it as:
Embeddings = Match meaning
Text = Match words
🧵🕵️
2
2
10
2,146
That was quick - @ZainHasan6 shows which Matryoshka sub-vectors text-embedding-3-large was likely trained on (512, 1024, 1536, 3072)
Can you see how much information the Matryoshka sub-vectors of the OpenAI text-embeddings-3-large model capture?
The graph shows the smoothed stdev. per dimension of 10k random samples from DBpedia embedded with the new text-emb3-large model.
We can see the variance is a multi-step function of the dimension index - earlier dimensions capture more information and those later on capture less.
The graph also tells us that the embedding model was trained with 4 aggregated loss functions at dim = {512d, 1024d, 1536d, 3072d} in the MRL style.
What's more interesting is that the information within the steps diffuses more or less evenly amongst the dimensions.
This also explains the magical finding in the MRL paper that shows you can actually take dimensions between these pre-specified chunks (eg. 750d) and still get benefits.
2
9
1,160
tl;dw
🏆 Compare embedding models using @huggingface’s MTEB leaderboard
💪 Different models perform better for different use cases (but there are also general purpose)
🔥 Run in Node.js & the browser using @xenovacom’s incredible transformers.js library
3
8
719
It all started with a small PR on their #Postgres repo [1], an invitation from @kiwicopple to write a blog post about it [2], and then a chance to help build their ChatGPT powered Supabase Clippy [3] 💚 📎
2
8
316
Thrilled to continue this work and dive into some really cool upcoming projects 🚀 🤫
(no, this wasn’t written by ChatGPT 🤖)
7
376
The whole experience opened my eyes to the quality (& stacked) team Supabase has brought together. I’m grateful to @kiwicopple, @AntWilson, and @saltcod for the opportunity to be a part of it 🙏
1
7
507
[1] github.com/supabase/postgres…
[2] supabase.com/blog/openai-emb…
[3] supabase.com/blog/chatgpt-su…
1
6
288
If you want the nitty-gritty's, feel free to join me down the rabbit hole:
piped.video/watch?v=Yhtjd7yG…
1
5
216
OpenAI back in the top 10 on the MTEB leaderboard via text-embedding-3-large.
Their support for custom dimension sizes to me is the game changer - I would happily trade 0.5 points on the MTEB board for 1/3 the vector size. Would love to know how this works under the hood.
3
1
6
503
Great stuff! 🎉 Let me know if you have any questions about how we do this @supabase
6
173
Background:
- Supabase uses PostgREST to power its auto-generated REST API (based on your DB tables)
- PostgREST supports a lot more SQL features than many people realize (joins, aggregates, etc)
- Sometimes its difficult to mentally translate your query to Supabase client code
- Use the translator to help make your life easier
1
6
686
Blog post: supabase.com/blog/database-b…
Open source: github.com/supabase-communit…
Shipped by: @_jgx_ 👏
1
6
2,359
We’re in the process of generating new benchmarks for HNSW at different compute addons (current benchmarks are for IVFFlat):
supabase.com/docs/guides/ai/…
Stay tuned!
cc @egor_test
1
5
270
Worth noting @andrewkane (the author of pgvector) mentioned he is working on product quantization which can improve speed & recall quite a bit.
4
179
Blog post is here:
supabase.com/blog/matryoshka…
We go into depth on Matryoshka embeddings, how embedding shortening works, and how you can implement Adaptive Retrieval in Postgres (with code)
4
585
This is a great question (updating the docs now). As long as your table structure matches what LangChain expects, you can instantiate the SupabaseVectorStore directly:
2
4
120
You definitely can 😄 example:
github.com/supabase-communit…
also supports streaming!
4
317
we need an @ollama-like Web API standard for in-browser LLMs
@xenovacom's epic phi-3 demo shows what's possible - the only missing piece is a way to cache and share models across domains (2GB is tiny, but I don't want to download that for every web app)
🤔
2
1
3
529
Thanks for your great work on MRL 🔥 I guess it's seen a bit of a revival recently 😄
yes lets chat!
1
4
154
Got it. Another approach that might work in the mean time is to create a view (or materialized view) that follows the document structure LangChain expects but fed dynamically with your trusted data.
1
3
92
Yep! The performance likely will not be the same as Pinecone, but you can get up and running right away and piggyback off of an existing Postgres DB. Worth looking into pgvector indexes once you do start scaling.
1
3
229
It’s a great question and something I haven’t had a chance to fully explore yet.
The scaling issues mostly comes down to indexes. pgvector offers one type of index today called IVFFlat. With lots of records it can be slow to build and take some trial-error to get right.
1
3
197
Currently Langchain vector stores & self query retrievers are pretty explicit about metadata, but there might be an opportunity to allow them to reference non-metadata columns for Supabase specifically. Were you hoping to filter over multiple joined tables?
1
2
121
This is a good combo - will try this soon. Doesn’t look like Ollama supports grammars yet, but hopefully soon:
github.com/jmorganca/ollama/…
1
1
41
We don't document this directly on Supabase, but we do on LangChain under their self-querying retrievers:
python.langchain.com/docs/in…
This converts a natural language query to a Supabase select query (via PostgREST under the hood), which is a bit safer than executing raw SQL.
2
60
You’re right! Looks like it’s composing Supabase CLI with Snaplet CLI results (both must have ranked high in embedding search). Thanks for the heads up! Likely we will filter out integration-related docs from the Clippy search and/or better prompt engineering.
2
89
Apart from fine tuning the model, explicit examples of writing style as part of your prompt could help. Also assuming you are performing RAG on your own content (containing your tone/style), you could encourage the model follow the same style as the injected RAG.
1
2
84
You got it - any time content changes you should generate a new embedding for that row. Thankfully this is trivial with triggers + edge functions - you can trigger these embedding to regenerate automatically when the row is updated, similar to what is shown in the video.
2
66
Thanks! This is something I'm hoping to improve very soon. In this case, the bad link is actually coming from OpenAI's pre-2021 training data of Supabase 🫠
Here's the same link sent directly from OpenAI's playground:
1
1
74
Check out @huggingface's MTEB leaderboard for models that perform just as good as text-embedding-ada-002 with 1/4 the dimensions:
huggingface.co/spaces/mteb/l…
2
180
Looks like you caught a bug with the translator 👀🙏 (joined table aliases causing issues)
Putting in a patch now, in the mean time can you try this?
1
2
74
Yes bge-small performs very well. The big thing to remember with bge-small is the need for a query instruction when using it for retrieval (search):
huggingface.co/BAAI/bge-smal…
2
61
Yes actually, but no need for a parameter. Are you using the supabase-js client or direct SQL? If js client, PostgREST actually supports filtering on top of rpc calls, which means you can make your WHERE as dynamic as you want by tacking on `.filter()` after `.rpc()`.
1
100
Also check out the SQL-to-REST translator to help convert your SQL queries to client lib / API requests (supports aggregates).
For the @supabase folks, introducing the SQL-to-REST API translator 🚀
Translate any SQL query to the equivalent PostgREST request or Supabase client code.
2
246
Just noticed this message at the bottom of your screenshot. Can you confirm whether you have the right permissions to update the org?
This message can probably stand to be more visible. Feel free to create a support ticket as well and we can help dive deeper into this.
1
28
We’ve investigated this, but sadly many vector databases make it difficult (or impossible) to fetch vectors in a way that would work with FDW.
If your embeddings are stored in another PG database though, this should be no problem.
1
54
Yes this looks like a hallucination. Good find - thanks for reporting! Yes feedback buttons are on the todo list!
2
38
You'll need a GPU with 6.4GB+ memory (works on Apple silicon, also tested on Windows).
Today you can run it on Chrome 113 (beta), or Chrome Canary.
1
2
189
Yes sir! github.com/supabase-communit…
1
2
28
You’ll still need to create the HNSW index yourself using that command (to give you most control), but it is available as of pgvector 0.5.0 which ships with all new Supabase databases (don’t forget to enable the vector extension first).
1
92
How @supabase's SQL-to-REST tool works under the hood
(if you have rabbit hole syndrome like me) 🐰🕳️🩺
2
143
Are your users looking up information based on specific keywords? Full text search.
Are they just looking for information that relates to their query, even if exact words don’t overlap? Consider embeddings!
2
2
225
That TS reference at the top should be pulling in the type. Looks like it might not be cached though based on the squiggly underlying. Can you try ctrl+./cmd+. on the url and cache it?
1
1
310
While you're at it, might as well also check out WebSD (stable diffusion in your browser)
github.com/mlc-ai/web-stable…
1
2
276