Is there anyone share the same feeling that langchain is over-designed?

[ludvik]

I put several hours trying to figure out my first demo and it seems you are basicly not able to use just part of it but have to use them all.

The classes in this library is entangled together by a big-tree-like multiple inheritance.

Building an APP from the ability of LLMs, Enbeddings should be clear and simple. This is not the way it should be.

[jjordanbaird]

This seems to me to be the first project of its kind to really catch on.
Lessons will be learned and future projects will have the luxury of seeing what worked/didn't work here.

[ludvik]

Messing around with the langchain code base for 3-4 days now, finally I am able to grasp some the design thoughts. I am now building my first App based on langchain.

Anyway I think this is a treasure to the Cognitive Application development community.

I can also see Harrison really immersing himself on this project, kudos for you! @hwchase17 and other contributors!

[tiagoefreitas]

Maybe someone should write a langchain agent that lets GPT-4 rewrite langchain from scratch (create the whole file structure) to improve it.
GPT-4 32K would have access to the whole repo and would have tools to create a new repo, create/merge PRs, test itself using test units, and run in a loop for some time.
Bonus if it can also use Claude and do it in turns (100 iterations GPT-4, 100 iterations Claude, repeat).
This could apply to any repo and GPT-4 could ask questions to a developer.

[linonetwo]

Then let AI use the new repo to improve it self, boom

[skiluo]

ye, I can't agree more

[linonetwo]

I'd suggest use BPMN, see https://github.com/bpmnServer/bpmn-server or https://camunda.com/ , it has visual editor and programming API, and have all workflow feature that langchain can provide.