ChatGPT plugins now support Postgres and Supabase | Hacker News Reader

ChatGPT plugins now support Postgres and Supabase

supabase.com

155 pointscendenta32 comments

I wrote a blog a few weeks ago on how LLMs will enable ‘Context Warehouses’ for work. Looks like they may be realised even sooner than I thought!

https://macroapp.io/blog/the-context-warehouse

I read your post but I'm not sure how async work relates to Context Warehousing. Maybe this was lost on me because I don't have context (hah) about your writing - maybe you often write about async work and this was a continuation of that. Isn't lack of context also an issue with sync teams?

Yes, my main point was that sync teams suffer from a lack of context, but that Context Warehouses will be able to help.

So a data lake of business artifacts queried using natural language and AI.

I think one disadvantage of pulling “context” is that you don’t really know what exists. You can get context for things you have an idea about, but not necessarily all the context you need.

Im not saying lots of meetings are necessarily great but you’re essentially indexing the information by listening.

Generally speaking though I like the idea of ingesting things as quickly as possible and then giving people strong tools to get what they need on demand.

hey hn, supabase engineer here

I just added a PR to the great retrieval plugin template repo from OpenAI to add Datastore options for Supabase and Postgres. This is an implementation of Datastore interface using Postgres database and 2 different interfaces: PostgREST for using with the ability to hook everything with your existing app and RLS policies, or just a plain database connection. They are completely swappable, the only thing you need to change is a couple of environment variables to switch from Postgres to Supabase and vice versa.

You will find an example of a plugin with full PostgresDB documentation embedded into postgres database in the post. And i will be happy to answer your questions here :)

Let me know if you want to play around with it and plug it in your ChatGPT, we can expose it if there is a need

Slightly tangent, but it would be really powerful to be able to train a model (through LoRa probably) on your datawarehouse, having table structures of facts and dimensions.

Self service BI would be a solved problem. No more building tedious widgets, but simply ask ChatGPT or similar for 'last 5 yours pNl by business unit'. Bonus points if based on cardinality, etc. it will properly choose the correct chart type (e.g.: stacked bar).

I’m building exactly this! https://vanna.ai

Awesome! Going to check it out

Your website says:

>Vanna.AI “trains” a model for your specific database

(Quotes mine)

When you say “trains”, I guess you don’t mean training in the true ML sense of updating model parameters, but rather you use a vector-db plus retrieval to augment queries (the usual recipe for “chat with my data”)?

We're building this at DinMo! (dinmo.com) We started with a context augmented assistant with retrieval and augmented queries, but we're actually working on training a smaller model for each client datawarehouse by creating a specific training dataset. You can check out a video the first version of the assistant here https://www.youtube.com/watch?v=qRClskxe1Wk

How is this considered private? When you use OpenAI's embedding model then aren't you already sending data to OpenAI?

This reads like Langchain implementation of "chat with your documents". Just that in this case instead of a standalone hosted chatbot, you do this from ChatGPT and can combine it with everything else ChatGPT might know.

is it the right convo branch? as this is a chatgpt plugin and uses openai embeddings api, it for certain sends some data to their servers. both when embeddings are generated and when chatgpt sends request to your plugin and receives response back.

Almost none of the "private LLM" ideas are private. There are hoards and hoards of idiots out there that continue to use OpenAI unfortunately.

The best thing to do is make your own Dockerfile to run Gaucamo locally. Putting it on gitlab too can help quench some of this cloud nonsense that keeps cropping up.

Slightly off topic: you see in the article one with a generic Engineering title (Egor) and one with fancy massive title CEO and Co-founder (Paul), while I’m almost certain that article and the work behind done by the engineering dude yet he’s not even given a proper title -at least of external marketing- while the other is piggybacking on that work. I saw that kind of behavior a lot especially in software companies, belittling, credit taking or sharing, etc., especially in small startups for some reason.

Hey, co-author here, thanks for caring about me. This is me who added myself as an author to the blog, i feel like “engineering” is the best description for the work I do in Supabase :) You can find it in PR https://github.com/supabase/supabase/pull/14520 We have a flat culture and if you scroll through past posts, you see that it is the most common description for our team.

As for the work behind, copple made significant edits and additions (as I thought of them) to my draft, so i decided add him as coauthor (just to make sure: I wasn’t asked by anyone and copple only noticed this after publication)

I don't know kiwicopple (apart from following Supabase), but I'm about 99% sure he's a dev and has probably built a ton of the features in Supabase. With a highly technical product like Supabase, you don't get to be a founder that plays with spreadsheets all day while devs build the product. You have to roll up your sleeves and build it.

In this particular situation, it simply doesn't fit whatever narrative you're weaving about some non-tech CEO taking the credit for the work done by a dev.

i'd spend a lot more time building if i had could!

Egor deserves all the credit for this announcement, but it was kind of him to include me for the small part I played. i'll consider removing my title for future posts to avoid any misconceptions

As a technical Founder, I completely feel ya :)

We have a joke in our company where I refer to the CEO-like tasks as "salad" and coding as "steak/dessert" – the salad is what keeps us healthy, sustainable, etc. But I'd love to have a lot more steak without affecting my "health". Even my director of operations flat out says "hey I need you to complete these salad tasks before you can enjoy your steak"

I was going to work on something similar, I'm glad I procrastinated.

You may still work on it. Not everyone’s daft enough to send customer data to openai and may need private models or alternative providers. Better yet, good alternatives to openai may free them up to focus on other things.

I wonder what would be the repercussions for Open AI. If propagandists feed it too much propaganda, or some unsuspecting federal employee uploads a classified document. Or some healthcare provider accidentally uploads confidential data. Would be interesting to see the reactions to it.

I have no doubt people are already uploading "just this time" bits of classified content and confidential data.

The propaganda thing, now that's an interesting idea.

Exactly, not everyone wants to send their data to OpenAI, I'm working on this (albeit very slowly) but hopefully I'll have something out in a week or two, and will be open source _and_ plug and play.

https://github.com/aldarisbm/memory

Does this use Faiss to find similar vectors?

https://engineering.fb.com/2017/03/29/data-infrastructure/fa...

heya, yep, if you add ivfflat index for embedding <column> in your table. I have to note that I am not pgvector contributor and this is how i understood it from the repo and its code. you can find a bit more context from the paper https://dl.acm.org/doi/pdf/10.1145/3318464.3386131 and https://github.com/pgvector/pgvector repo itself

Dumb question how do I get chatgpt plugins? I have a paid account and asked for it but don’t seem to have it, and it seems like everyone else does.

You'll need to request access here: https://openai.com/waitlist/plugins

Self-update I figured it out, it’s available if you have a paid plan but a couple levels deep in the settings menu buried under beta features tag.

Who gives a shit

Reply on news.ycombinator.com