Pass arguments to Langgraph node

[nathan-vo810]

I have a question regarding how to pass additional arguments when running the graph. For example, we have the following snippet of code:

def fn(state, extra_val_1, extra_val_2):
    do_sth(extra_val_1)
    do_sth(extra_val_2)

...
workflow.add_node("fn", fn)
...
app = workflow.compile()
app.invoke(input)

So, how can I pass values for extra_val_1 and extra_val_2?

[hinthornw]

To pass things to a node, there's really two easy ways:

1. In the state: if you don't want it used by other nodes, namespace it:

class State(TypedDict):
    messages: list
    my_super_secret_args_for_fn: dict

2. as configurable params

def my_node(state, config):
   some_arg = config["configurable"]["extra_val_1"]

....
app.invoke(input, {"configurable": {"extra_val_1": 42}})

[ShonenAsh]

If the node function sits in a different python file/class, how to pass an LLM and a VectorStoreRetriever instance to the node when adding the node function to the graph?

Passing it through config doesn't feel right.

[hamdan-27]

@ShonenAsh I faced the same issue as you, I had to redefine the LLM and VectorStoreRetriever in each node function separately for it to work

[ShonenAsh]

@ShonenAsh I faced the same issue as you, I had to redefine the LLM and VectorStoreRetriever in each node function separately for it to work

Well I changed my approach entirely so that I dont have to pass LLM/VectorStore objects to every Node (and you should too).

Put your Langgraph nodes (functions) in a class and make the LLM and other such variables part of the class's object.

Then you can self.llm anywhere in your functions

[shadowtechn]

@ShonenAsh It does work, and I use it as my default design. However, I’ve recently started using LangGraph Studio, and it requires a module-level variable that holds the compiled graph. That doesn’t seem to work when the graph is defined inside a class.

How do you handle this?

[hectoradrian961030]

I faced a similar situation where I needed to pass an LLM to the state. I solved it by using partial:

llm = ChatOpenAI(...)
builder.add_node("process_confirmation", partial(process_confirmation, llm=llm))

and the function

def process_confirmation(
state: ConversationState,
llm: ChatOpenAI
) -> ConversationState:
"""Process the user's confirmation"""
structured_llm = llm.with_structured_output(ProcessConfirmation)

[87royalts87]

I faced a similar situation where I needed to pass an LLM to the state. I solved it by using partial:

llm = ChatOpenAI(...) builder.add_node("process_confirmation", partial(process_confirmation, llm=llm))

and the function

def process_confirmation( state: ConversationState, llm: ChatOpenAI ) -> ConversationState: """Process the user's confirmation""" structured_llm = llm.with_structured_output(ProcessConfirmation)

Thanks for the reply :)
By using the functools.partial my graph breaks. (e.g. graph.get_graph().draw_mermaid_png()) shows rubbish. I stil wonder what is the pythonic way to pass extra arguments to the nodes... 🙈

[danikenan]

builder.add_node("position_refined", position, metadata={"__my_": "his"})
builder.add_node("position", position, metadata={"__my__": "yours"})
....
def position(state: State, config: RunnableConfig):
  my = config["metadata"]["__my__"]