I have a similar feeling looming at the converse API.
Why, do we have a thing that keeps returning all the text, when something on the other end is just appending to it.
Then there is the code inside langchain, some of which feels rushed.
I'm unconvinced of the abstraction that everything is a "Document".
In an app I'm working on, once we switched to PGVector in Django the need for a lot of things went away.
What is great with langchain is the support for lots of things.. but not everything.
So, wanting to always use a thin abstraction over native libraries we find ourself using litellm, which covers some bits and langchain for the others (though the code for both of those is not much).
And then there's models: we can't even agree on standard names for the same models.
And this makes it painful when you support different backends, of course if Bedrock is a backend they have their own models you can't use anywhere else.
I have a similar feeling looming at the converse API.
Why, do we have a thing that keeps returning all the text, when something on the other end is just appending to it.
Then there is the code inside langchain, some of which feels rushed.
I'm unconvinced of the abstraction that everything is a "Document". In an app I'm working on, once we switched to PGVector in Django the need for a lot of things went away.
What is great with langchain is the support for lots of things.. but not everything.
So, wanting to always use a thin abstraction over native libraries we find ourself using litellm, which covers some bits and langchain for the others (though the code for both of those is not much).
And then there's models: we can't even agree on standard names for the same models.
And this makes it painful when you support different backends, of course if Bedrock is a backend they have their own models you can't use anywhere else.