LangChain and LangGraph

pip install 'agent-once[langchain]'

once.integrations.langchain.once_tool wraps a function for a LangChain tool. Put @tool on top:

from langchain_core.runnables import RunnableConfig
from langchain_core.tools import tool

from once import Once
from once.integrations.langchain import once_tool

once = Once("once.db")


@tool
@once_tool(once, name="send_invoice")
def send_invoice(customer: str, amount: int, config: RunnableConfig) -> str:
    """Send an invoice to a customer."""
    return billing.send(customer, amount)


graph.invoke(inputs, config={"configurable": {"thread_id": "order_456"}})

once_tool(
    once,
    fn=None,
    /,
    *,
    name=None,
    timeout=None,
    ttl=None,
    verify=None,
    scope_key="thread_id",
    config_arg=None,
)

name, timeout, ttl and verify work as in @once.tool; ttl is seconds or a timedelta (see Expiry and gc).

The scope is, in order:

1. config["configurable"][scope_key], the LangGraph thread id by default.
2. the enclosing once.scope(), if any.

The config argument is found by its RunnableConfig annotation or the name config; config_arg= names it explicitly. LangChain fills it in and hides it from the model. It is left out of the key.

When a LangGraph thread resumes from a checkpoint and replays a node, the tool calls it already made return their stored results instead of running again.

Sync and async functions both work. examples/langgraph_example.py in the repo is a runnable graph.