Add support for Supabase VectorStore

[davidmigloz]

supabase/supabase-flutter#489 (comment)

https://cookbook.openai.com/examples/vector_databases/supabase/readme

[matteodg]

Hi @davidmigloz , I'd like to take care of this task, is that OK?

[davidmigloz]

Hey @matteodg !

Thanks for your interest in contributing 🙂

@luisredondo was going to start working on it (not sure if he started implementing it already), but I see you already made quite some progress so go for it!

[davidmigloz]

Actually, @luisredondo is telling me that he already started implementing it as well (my bad, I didn't update the ticket). It'd be good to try to sync/merge your changes and his changes. Are you in the LangChain.dart Discord @matteodg? We can discuss it in the #contributing channel.

[luisredondo]

Hey @matteodg! I have been researching how to implement this appropriately and can offer some input if needed. We can sync via Discord. @davidmigloz provided the link above. You already made a good amount of progress, so I don't mind if you take over it. But here I am if you need a hand!

[matteodg]

Hi @luisredondo! I'm basically converting to Dart from the Python LangChain (but actually the JS one might be a better source, easier for me at least as I'm more a Java dev). Let me know any input you have or any work you already have done. Thanks!

[matteodg]

I found a few inconsistencies between Python LangChain and the JS one:

• Python Langchain assumes the documents table has an id of type UUID, see https://python.langchain.com/docs/templates/self-query-supabase#setup-supabase-database
• JS LangChain assume the documents table has an id of type BIGSERIAL and has the match_count in the match_documents function, see https://js.langchain.com/docs/integrations/vectorstores/supabase

I used the JS LangChain definition for the moment.

[matteodg]

Hi @davidmigloz and @luisredondo, the PR is ready for you to review!

I slightly changed the match_documents functions: it now contains also the match_threshold (see the blog entry Storing OpenAI embeddings in Postgres with pgvector), so we can filter the results by threshold directly in Supabase instead of filtering them locally.

[davidmigloz]

@matteodg awesome, thank you very much!

Can you open a PR from your repo to this repo?
I'll review it tomorrow

[matteodg]

Here it is #318

[davidmigloz]

Released 🚀
https://pub.dev/packages/langchain_supabase/changelog#001

Docs: https://langchaindart.com/#/modules/retrieval/vector_stores/integrations/supabase