Skip to main content

Parallel Execution and Branching

Real chains are rarely a single straight line. RunnableParallel runs several steps against the same input concurrently, RunnablePassthrough threads the original input through unchanged so a later step can still see it, and RunnableBranch picks one path out of several based on a condition.

RunnableParallel — fan-out​

from langchain_core.runnables import RunnableParallel

chain = RunnableParallel(
    summary=summarize_chain,
    sentiment=sentiment_chain,
)
chain.invoke({"text": "..."})
# {"summary": "...", "sentiment": "..."}
tip

The dict literal {"summary": summarize_chain, "sentiment": sentiment_chain} is a RunnableParallel — LCEL coerces it automatically. That's the idiom you'll see everywhere in RAG examples, where a chain fans out to {"context": retriever, "question": RunnablePassthrough()}.

RunnablePassthrough — keep the original input​

Once a chain transforms its input, the original is gone unless something preserves it. RunnablePassthrough is an identity Runnable — it just returns whatever it's given — used inside a RunnableParallel to carry the original input alongside a derived one.

from langchain_core.runnables import RunnablePassthrough

chain = RunnableParallel(
    context=retriever,
    question=RunnablePassthrough(),
) | prompt | model | parser

RunnableBranch — conditional routing​

from langchain_core.runnables import RunnableBranch

branch = RunnableBranch(
    (lambda x: "code" in x["topic"], code_chain),
    (lambda x: "billing" in x["topic"], billing_chain),
    general_chain,  # default, no condition
)

Each tuple is (condition, runnable); the first condition that returns truthy wins. The final positional argument (no tuple) is the default.

See also​