Avoid Reloading Model on Each Agent Call in LangGraph While Using Configurations

[farouk09]

In LangGraph, is it possible to use configuration parameters (like model_name or system messages) at the graph level rather than only inside a specific function like call_model? My issue with the current approach is that every time we call the agent, the call_model function is triggered and the model gets loaded again, which is time-consuming. I’d like to load the model once outside the function and still use configuration — is that possible?

def call_model(state, config):
    messages = state["messages"]
    model_name = config.get('configurable', {}).get("model_name", "anthropic")
    model = _get_model(model_name)
    response = model.invoke(messages)
    # We return a list, because this will get added to the existing list
    return {"messages": [response]}

[hinthornw]

Ya: https://python.langchain.com/docs/how_to/chat_models_universal_init/#creating-a-configurable-model

model = init_chat_model()
def my_node(state, config):
   model.invoke("foo", config)