langgraph/how-tos/command/

[giscus[bot]]

langgraph/how-tos/command/

Build language agents as graphs

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

[DorSimoni]

Hey,
First, thank you for your contribution and efforts.

I'm encountering some issues with implementing "Command".
I'm using the most updated LangGraph version (0.2.56).
I see that "goto" parameter does not exist in Command class, and "update" parameter does not exist in GraphCommand class either.
Did I miss something?

My goal is to iterate through a list of objects. For each object that I classify as "yes," I want to return it to the user (edge END?) while, at the same time, allowing the graph to continue classifying the remaining objects.

Within the loop, I have conditions to check whether I have classified enough objects and whether I have classified enough "yes" responses.
I thought using Command might be a good approach for this.
Would you suggest a different solution?

Thanks,
Dor

[vbarda]

@DorSimoni you shouldn't see GraphCommand at all in the latest version of LangGraph, only Command, could you please double check in a fresh virtual environment?

as for you idea, could you clarity: what does it mean to "return to the user" in this case? is the user supposed to take any additional actions?

[DorSimoni]

Thanks for your quick response.
I tried it in a clean virtual environment, and the Command function now has the correct parameters.

As for the idea, sorry for the confusion. By "return to the user," I meant that the graph will output an answer to the user while continuing to classify more documents. Each document classified as "yes" will be output to the user.
The user is not supposed to take any additional actions (at least not at this stage. in the next stage, they will, but that's a different challenge for HILP).

[samgoos7]

I am having issues with using Command, im getting the following error: TypeError: Command.init() got an unexpected keyword argument 'goto'
Even when running the exact examples in shown on this page, including the imports.

Am I doing something wrong?

Thanks,
Sam

[vbarda]

hm, i would recommend reinstalling in a fresh virtualenv, but please let me know if that doesn't help

[dillon-nidoai]

Is this meant to replace SEND? Or be more of an augmentation to be able to map reduce while updating state. Should commands replace conditional edges?

[dillon-nidoai]

goto: Union[Send, Sequence[Union[Send, str]], str] = ()

It seems you can also use Send within command. So could you map reduce with it as well? I guess I'm wondering if you can map send with command and continue with different nodes of different branches. not sure if there would be individual states though.

[gbaian10]

@samgoos7 You mentioned a question I've always wanted to know about.
In the past, when there was only a conditional edge, were we unable to both enter the edge's route and update the state simultaneously?

Was the Command introduced to solve this issue?

[dillon-nidoai]

I don't think so. Since state updates only when you return the state from my experience. So you would have to update state before entering a conditional edge then return the node you want to route to.

I was more wondering if you could branch and run branches in parallel with independent instances of state with Command. For example, I was playing with a conditional [Send("update_node") for s in state["value"]] to another conditional (update_node, router) which returned the node to go to. But I wasn't able to continue down these branches and would have to reduce. So I wonder with Command if I can maintain and run those independent instances then reduce later. (Sorry I know this wasn't really part of the question lol)

[samgoos7]

@gbaian10 yeah at least that is my use case. For example retrieving tool content message from messages state and updating the state, and go to the next node based on that update. Previously you had to do that with different steps (update state node + router node for conditional edge) and now you can do this simoultaneously with Command

[vbarda]

Command was primarily introduced to allow combining control flow (edges) and state updates (which as folks pointed out was a common problem with conditional edges). it is NOT meant to replace Send but you can use it with Send as well, we'll add examples

[Saisiva123]

I think here the output should be {'foo': 'ab'} right? because inside the node A the graph state will get updated to {'foo': 'a'} and then it will navigate to b where the state gets updated to {'foo': 'ab'}.

[Saisiva123]

Got confused because inside the node A, it would have been better if we have choosen 'b' and 'c'. random.choice(["b", "c"]).

But the output for the above code should look something like this.
Called A
{'node_a': {'foo': 'a'}}
Called B
{'node_b': {'foo': 'ab'}}

---

Called A
{'node_a': {'foo': 'b'}}
Called C
{'node_c': {'foo': 'bc'}}

[sctrmn]

How a ToolNode should use Command?

[Simon-U]

I tried it, and it seems to work just fine with tools using the Command

[Blaqadonis]

Could you do a tutorial for an assistant that prompt's a user's feedback many times in the workflow and at different nodes?

[Blaqadonis]

Nodes b and c can also be written like this right?
def node_b(state: State):
print("Called B")
return Command(
update={"foo": state["foo"] + "b"}
)

def node_c(state: State):
print("Called C")
return Command(
update={"foo": state["foo"] + "c"}
)

[ajeeth-b]

Unable to See Edges with Class-Based Nodes using Command and get see edges in graph.get_graph(xray=True)

I was trying a class based node, but unable to see edges when I do graph.get_graph(xray=True)

this is my class based node definition

class MyNode(AbastractClassWithRunMethodOnly)
    def run(self, state, *args) -> Command[Literal['node2', END]]:
         # my logic
         return Command(goto='node2')

# has graph obj
my_node_obj = MyNode()
graph.add_node(my_node_obj.run)

this way I'm not able to see the edges .
is there any work around or I'm doing something wrong?

[ajeeth-b]

resolved above issue in this release few hours ago

https://github.com/langchain-ai/langgraph/releases/tag/0.2.63