Is it possible to hide graph state variables from langsmith logs?

[thoffmann-artidis]

Hi,

I am working on a multi-agent chat application where specific agents require user authentication to be passed to the graph. I.e. something like that:

from typing import TypedDict, Literal

from langchain_core.runnables import RunnableLambda
from langchain_openai import ChatOpenAI
from langchain.prompts import ChatPromptTemplate
from langgraph.graph import StateGraph


llm = ChatOpenAI(model='gpt-4o')


class State(TypedDict):
    message: str
    bucket: Literal['cars', 'animals']
    user_secret_key: str
    answer: str


def get_user_buckets(state: State) -> dict[str, str]:
    ans = ''
    if state['user_secret_key'] == 'user_1':
        if state['bucket'] == 'cars':
            ans = 'Toyota, Audi, Mercedes'
        elif state['bucket'] == 'animals':
            ans = 'horse, cat'
    elif state['user_secret_key'] == 'user_2':
        if state['bucket'] == 'cars':
            ans = 'Opel, Bugatti'
        elif state['bucket'] == 'animals':
            ans = 'bird, dog, fish'
    return {'answer': ans}


prompt = ChatPromptTemplate.from_messages(
        [('system', ('You are supposed to select a bucket based on the user'
                     ' request by setting the JSON `{{"bucket": ...}}`.')),
         ('user', '{message}')])

supervisor = prompt | llm.with_structured_output(State, method="json_mode")
agent = RunnableLambda(get_user_buckets)

workflow = StateGraph(State)
workflow.add_node('supervisor', supervisor)
workflow.add_node('agent', agent)
workflow.add_edge('supervisor', 'agent')
workflow.set_entry_point('supervisor')
graph = workflow.compile()

graph.invoke({'message': 'Show me all cars in the bucket.',
              'user_secret_key': 'user_1'})

Given that I'm using langsmith to record the traces, is there a neat way to hide the state variable user_secret_key from the langsmith logs? If not, is there a better way to pass secrets to agents/tools during runtime?

Thanks!

[thoffmann-artidis]

Solution: https://python.langchain.com/docs/how_to/runnable_runtime_secrets/