LangChain SDK Patterns (Python)

LangChain SDK Patterns (Python)

Overview

chain.batch(inputs) in LangChain 1.0 does not parallelize by default. The max_concurrency parameter defaults to 1 in several provider packages (notably older langchain-openai), so a call like chain.batch(inputs_1000) runs 1,000 sequential round-trips — same wall-clock time as a for loop, plus the overhead of the batch machinery. Users file "batch is slow" tickets, benchmark it against asyncio, and move to a different framework — when the fix is two lines:

# BAD — silently serializes (P08)
chain.batch(inputs_1000)

# GOOD — 10 in flight at once
chain.batch(inputs_1000, config={"max_concurrency": 10})

Then three more traps wait:

• P07 — .with_fallbacks([backup]) defaults exceptions_to_handle=(Exception,), and on Python <3.12 that tuple includes KeyboardInterrupt. A Ctrl+C during a long run does not stop the process — it silently hands off to the fallback chain and keeps billing.
• P57 — ChatPromptTemplate.from_messages(..., template_format="f-string") (the default) parses every { in every string, including user input. A user who pastes {"error": "..."} raises KeyError: 'error' at invoke time.
• P53 — Pydantic v2 rejects extra fields by default; models cheerfully add summary or confidence to your Plan schema and with_structured_output crashes with ValidationError: extra fields not permitted.

This skill walks through LCEL composition (RunnableSequence, RunnableParallel, RunnableBranch, RunnablePassthrough, RunnableLambda); the correct exceptions_to_handle whitelist per provider; max_concurrency tuning with safe ceilings (10 for most providers, 20+ with a semaphore); and prompt templates that survive untrusted input. Pin: langchain-core 1.0.x, langchain-anthropic 1.0.x, langchain-openai 1.0.x. Pain-catalog anchors: P07, P08, P53, P57.