LangChain Core Workflow (Python)
Overview
An engineer wires a four-stage LCEL chain: classify the question, retrieve
context, format the prompt, invoke the LLM. It looks clean:
chain = (
RunnablePassthrough.assign(category=classifier)
| RunnablePassthrough.assign(docs=retriever)
| prompt
| llm
| StrOutputParser()
)
chain.invoke({"question": "What's our refund policy?"})
The call returns this:
Traceback (most recent call last):
...
File ".../runnables/base.py", line 3421, in _call_with_config
output = call_func_with_variable_args(func, input, ...)
File ".../prompts/chat.py", line 1021, in _format_messages
return await ... await self.ainvoke({**kwargs})
KeyError: 'question'
Nothing in that stack says which stage produced the wrong dict shape. The
RunnablePassthrough.assign(docs=retriever) call silently rebuilt the dict
and — because retriever was itself a Runnable[str, list[Document]] that
took the question string, not the dict — a mis-piped intermediate value
overwrote the question key. The prompt template expected {question} and
blew up. This is P06 in the pack's pain catalog: .pipe() on mismatched dict
shape raises KeyError deep in runnable internals with no hint at the
offending stage.
The fix is two patterns you install once and never remove:
- Debug probes — a
RunnableLambda that logs dict keys between every two
stages. <1ms overhead per invocation. Surfaces the exact stage that mutates
the shape.
- Typed composition — annotate each chain with
RunnableSerializable[InputT, OutputT]
plus pydantic BaseModel types so mypy flags the mismatch at lint time
instead of at .invoke().
Meanwhile, a second trap waits for anyone tempted to wrap tool-using chains
in the legacy AgentExecutor: it silently swallows intermediate tool errors
as empty-string observations and the agent cheerfully answers "I couldn't
find the answer" (P09). For agent loops in LangChain 1.0, skip AgentExecutor
and use LangGraph's create_react_agent — errors raise, not vanish. This
skill cross-references langchain-langgraph-agents (L26) for that path.
Composition primitives covered — with input/output shapes and use cases — are
RunnableParallel (fan-out, 2–3× wall-clock win on 2 independent retrievals),
RunnableBranch (conditional routing with mandatory default), RunnablePassthrough.assign
(merge computed fields without losing input), and RunnableLambda (arbitrary
Python, used here for debug probes and shape assertions). Pin: langchain-core 1.0.x.
Pain-catalog anchors: P06, P09.
Prerequisites
- Python 3.10+
langchain-core >= 1.0, < 2.0pydantic >= 2.0 for typed composition- At least one chat provider installed (see
langchain-model-inference)
- Familiarity with the composition primitives introduced in
langchain-sdk-patterns
Instructions
Step 1 — Fan-out with RunnableParallel for independent sub-tasks
RunnableParallel({"a": chain_a, "b": chain_b}) forwards the same input to
both branches and runs them concurrently, merging results into a dict.
Wall-clock collapses to the slower branch — 2–3× speedup on two parallel
retrievals is typical (dense vector search + BM25, or tool-use + analysis).
from langchain_core.runnables import RunnableParallel, RunnablePassthrough
hybrid = RunnableParallel(
dense=dense_retriever,
bm25=bm25_retriever,
query=RunnablePassthrough()
)
result = hybrid.invoke("refund policy for damaged goods")
Input shape: whatever the sub-chains accept (all must accept the same shape).
Output shape: dict with one key per sub-chain. If one branch takes 5× longer,
your total latency is that branch — not the sum. See
Parallel vs Sequential for the async
.abatch() variant, shared-state gotchas, and a benchmarking template.
Step 2 — Route on input with RunnableBranch
RunnableBranch((cond, runnable), ..., default) dispatches per-input to the
first matching branch. The default is mandatory — without one you get a
silent fallthrough on unmatched inputs.
from langchain_core.runnables import RunnableBranch
router = RunnableBranch(
(lambda x: x["category"] == "refund", refund_chain),
(lambda x: x["category"] == "shipping", shipping_chain),
(lambda x: len(x["question"]) > 2000, long_form_chain),
general_chain,
)
Classifier-gated routes are the common case: a cheap small-model classifier
runs first, its label goes into the dict via RunnablePassthrough.assign, and
RunnableBranch dispatches. See Branch Routing Patterns
for the signature, classifier-gated route recipe, fallback route pattern, and
pytest patterns for each branch in isolation.
Step 3 — Thread state with RunnablePassthrough.assign
.assign(field=...) merges a computed field into the input dict without
losing any existing keys. This is the primary pattern for staged context
assembly in RAG:
from langchain_core.runnables import RunnablePassthrough
def format_docs(docs: list) -> str:
return "\n\n".join(d.page_content for d in docs)
staged = (
RunnablePassthrough.assign(docs=retriever)
| RunnablePassthrough.assign(context=lambda x: format_docs(x["docs"]))
)
staged.invoke({"question": "..."})
The input dict passes through unchanged; the new field is the only mutation.
This is the safest shape-preserving primitive in LCEL — use it whenever a
downstream stage needs both the original input and a computed value. See
Passthrough Assign Patterns for
staged context assembly, the itemgetter variant for pulling a single field
into a typed chain, and anti-patterns that re-shadow input keys.
Step 4 — Use RunnableLambda for debug probes (and little else)
RunnableLambda(fn) wraps any Python callable into a runnable. Its best use
is debug probes — log intermediate values without breaking the pipe:
from langchain_core.runnables import RunnableLambda
def probe(stage: str):
"""<1ms overhead per invocation. Returns input unchanged."""
def _probe(x):
keys = list(x.keys()) if isinstance(x, dict) else type(x).__name__
print(f"[probe:{stage}] keys={keys}")
return x
return RunnableLambda(_probe)
chain = (
probe("input")
| RunnablePassthrough.assign(category=classifier)
| probe("after-classify")
| RunnablePassthrough.assign(docs=retriever)
| probe("after-retrieve")
| prompt
| probe("after-prompt")
| llm
| StrOutputParser()
)
When P06's KeyError strikes, the last probe that printed tells you exactly
which stage produced the wrong shape. Remove the probes (or gate them on an
env var) after debugging. For production chains that stay observable, prefer
langchain.debug = True or LangSmith tracing over print probes.
See Debug Probes for a shape-assertion decorator,
the langchain.debug flag, verbose mode, and a probe that raises instead of
prints (useful in CI).
Avoid RunnableLambda for real logic — it loses LangSmith tracing fidelity
(input/output become opaque blobs) and a >3-line lambda is a sign you want a
concrete Runnable subclass. See the anti-pattern note in
langchain-sdk-patterns/references/runnable-composition-matrix.md.
Step 5 — Type chains with RunnableSerializable[InputT, OutputT]
The root fix for P06 is static typing at chain boundaries. RunnableSerializable
carries input and output type parameters; pair it with pydantic BaseModels
and mypy catches dict-shape mismatches at lint time:
from pydantic import BaseModel
from langchain_core.runnables import RunnableSerializable
class RAGInput(BaseModel):
question: str
user_id: str
class RAGOutput(BaseModel):
answer: str
citations: list[str]
def build_rag_chain() -> RunnableSerializable[RAGInput, RAGOutput]:
return (
RunnablePassthrough.assign(docs=retriever)
| RunnablePassthrough.assign(context=lambda x: format_docs(x["docs"]))
| prompt
| llm.with_structured_output(RAGOutput)
)
rag: RunnableSerializable[RAGInput, RAGOutput] = build_rag_chain()
Call rag.input_schema.model_json_schema() to dump the runtime-enforced
schema for tests or assertions. Combined with the debug probes from Step 4,
this is the typed-composition pattern that retires P06 in new code.
Step 6 — RAG composition example (end-to-end)
from langchain_core.prompts import ChatPromptTemplate
from langchain_core.output_parsers import StrOutputParser
from langchain_core.runnables import RunnablePassthrough, RunnableParallel
prompt = ChatPromptTemplate.from_messages([
("system", "Answer from the context. If not in context, say so."),
("human", "Context:\n{{ context }}\n\nQuestion: {{ question }}"),
], template_format="jinja2")
rag_chain = (
RunnableParallel(
dense=dense_retriever,
bm25=bm25_retriever,
question=RunnablePassthrough(),
)
| RunnablePassthrough.assign(
context=lambda x: format_hybrid(x["dense"], x["bm25"])
)
| prompt
| llm
| StrOutputParser()
)
answer = rag_chain.invoke("What's the return window for electronics?")
This is the reference composition — parallel dual-retrieval (2–3× speedup vs
sequential), staged context assembly via .assign, jinja2 prompt templating
(escapes literal { from retrieved docs, see langchain-sdk-patterns), and
a simple string output. Add debug probes from Step 4 during development.
Step 7 — For agent loops, skip AgentExecutor
Legacy AgentExecutor silently catches tool exceptions and feeds the error
message back as an empty-string observation — the agent then answers "I
couldn't find the answer" with no trace of the underlying failure (P09). In
LangChain 1.0, use LangGraph's create_react_agent instead: tool errors
raise by default, intermediate steps are inspectable, and recursion limits
are explicit.
See the pack's langchain-langgraph-agents skill (L26) for the agent
migration path. The composition primitives in this skill (RunnableParallel,
RunnableBranch, .assign) remain the right tools for the non-agent parts
of any LangGraph workflow — they compose inside a LangGraph node.
Composition Pattern Table
| Pattern | Input shape | Output shape | Typical use |
|---|
a | b | c (RunnableSequence) | a.input | c.output | Linear pipeline: prompt → llm → parser |
RunnableParallel(x=ch1, y=ch2) | forwarded to both | {"x": ch1.output, "y": ch2.output} | Fan-out: hybrid retrieval, tool-use + analysis |
RunnableBranch((cond, ch), ..., default) | anything cond accepts | ch.output or default.output | Classifier-gated routing, per-input dispatch |
RunnablePassthrough() | any | same as input | Keep original value alongside a transform in a parallel |
RunnablePassthrough.assign(k=ch) | dict | input dict with "k" added (ch.output) | Staged context assembly, threading state |
RunnableLambda(fn) | fn's arg | fn's return | Debug probes, shape assertions — avoid for real logic |
Output
- Multi-step chains composed from
RunnableParallel, RunnableBranch, RunnablePassthrough.assign, and RunnableLambda with declared input/output shapes
- Debug-probe pattern (<1ms per invocation) that surfaces the exact stage producing a wrong dict shape when P06 strikes
- Typed-composition pattern (
RunnableSerializable[InputT, OutputT] + pydantic BaseModel) that catches P06 at lint time before invocation
- RAG composition recipe: hybrid parallel retrieval + staged context assembly + prompt + llm + parser, 2–3× wall-clock win on dual retrieval
- Cross-reference to
langchain-langgraph-agents (L26) for agent loops that avoid P09's silent-error trap
Error Handling
| Error | Cause | Fix |
|---|
KeyError: 'question' deep in runnable internals, no stage named in stack | RunnablePassthrough.assign or an upstream runnable rebuilt the dict in a way a downstream stage does not expect (P06) | Insert RunnableLambda debug probes between stages; annotate with RunnableSerializable[InputT, OutputT] + pydantic for lint-time catch |
| Agent returns "I couldn't find the answer" with no error in logs | Legacy AgentExecutor swallows tool exceptions as empty-string observations (P09) | Migrate to LangGraph create_react_agent; see langchain-langgraph-agents (L26) |
RunnableBranch returns wrong branch's output for unmatched input | No default branch supplied; LCEL fell through to the last-declared pair | Always pass a default as the final positional arg to RunnableBranch((cond, ch), ..., default_ch) |
.assign computed field shadows input key | RunnablePassthrough.assign(question=...) overwrote the original question | Name computed fields distinctly from input keys; use itemgetter to extract when you need the original |
RunnableLambda call shows as opaque blob in LangSmith | Lambda's input/output are not traced at block-level | Replace with a concrete Runnable subclass or use .assign(key=...) with a named function |
RunnableParallel branches race on shared mutable state | A lambda closed over a non-thread-safe object | Keep parallel branches side-effect-free; materialize shared state before the parallel block |
Examples
Hybrid retrieval with parallel branches
Two retrievers run concurrently on the same query — dense vector search via
an embedding model and BM25 via rank-bm25. RunnableParallel merges the
results; a downstream .assign step formats the union. Wall-clock is the
slower retriever, not the sum — typically a 2–3× win over sequential dense
then BM25.
See Parallel vs Sequential for the
benchmarking template and async variant.
Classifier-gated routing for a customer support chain
A cheap small-model classifier runs first and tags the input dict with a
category. RunnableBranch dispatches refunds, shipping, and general
questions to specialist chains. The default branch handles unmatched inputs
with a disclaimer prompt.
See Branch Routing Patterns for the
classifier-gated recipe and pytest isolation patterns.
Staged context assembly for RAG with citations
RunnablePassthrough.assign(docs=retriever) then
.assign(context=format_docs) builds the prompt context without dropping the
original question. A final .assign(citations=lambda x: [d.metadata["url"] for d in x["docs"]])
threads citations through to the structured output.
See Passthrough Assign Patterns
for the staged pattern and the anti-pattern around key shadowing.
Debug probe that raises in CI
In development, probes print. In CI, they assert on expected keys and raise
on mismatch — catching P06 during test runs instead of at production invoke.
See Debug Probes for the shape-assertion
decorator and the langchain.debug flag.
Resources
