Forking the graph within the graph workflow

[CRSiegfried]

I have a scenario where a node is reached and sometimes the best step forward for that node is to generate a set of parallel teams that have the full history of the current graph state with some select values altered. It seems like the most direct route to do this would be to utilize checkpointing, but I'm not sure about the cleanest way to handle calling the graph (and its checkpoint states) from within the graph. I would rather not generate a super-graph to manage this.

I'm hoping I'm just missing a step and there is a straight-forward way to fork the graph from within its workflow and appreciate any thoughts.

[hinthornw]

Hi @CRSiegfried - I believe the Send API would work for this?

https://langchain-ai.github.io/langgraph/how-tos/map-reduce/

LMK if you run into issues with it. The basic idea is to support branching with custom copies/views of the state (so you can use a clone of a node /team and apply it to a different value)

[hwchase17]

echoing this answer, if you want branching to part of the "cognitive architecture" of the application, i would put that logic as part of the graph (using Send should work well). checkpointing should be used for human introspection/management of the graph

[CRSiegfried]

Send does fit the bill, thank you!

I was using a weakref registry with custom serialization so I could pass the compiled graph via the StateGraph. That still method still has some interesting possibilities, but send is a lot cleaner and fits most of the use-cases I was considering.