How do you save and load a LangChain chain for reuse?

What you need to know

"Saving a chain" means three different things, and each has its own right answer.

The recommended pattern: a factory function

1from langchain.chat_models import init_chat_model2from langchain_core.prompts import ChatPromptTemplate3from langchain_core.output_parsers import StrOutputParser45def build_summary_chain(cfg: dict):6    prompt = ChatPromptTemplate.from_messages([7        ("system", cfg["system_prompt"]),       # loaded from a YAML file8        ("human", "{text}"),9    ])10    llm = init_chat_model(cfg["model"], temperature=cfg.get("temperature", 0))11    return prompt | llm | StrOutputParser()

The code is versioned in Git and unit-tested; the configuration can differ per environment. This is how LangChain apps are normally shipped.

When you do need serialisation

1from langchain_core.load import dumps, loads23blob = dumps(prompt, pretty=True)          # JSON text, safe to store4same_prompt = loads(blob)                  # core classes are allowed by default

Key facts, checked against langchain-core 1.6:

• Only classes marked serialisable round-trip. A RunnableLambda holding your own Python function does not.
• Secrets are replaced by placeholders, so the API key is never written out.
• By default loads only allows core classes. A chain containing ChatOpenAI needs an explicit allowlist, and secrets are only read from the environment if you pass secrets_from_env=True.
• A serialised payload is executable configuration: it can set a base_url or headers. Load only what you wrote yourself.
• prompt.save() and load_prompt() are deprecated in favour of dumps and loads.

A real-life example

A lead-qualification pipeline for a Hyderabad real-estate firm scores website enquiries as hot, warm or cold. The first version pickled the whole chain to S3 so the worker could load it. After a library upgrade, the pickle no longer loaded and the worker crashed at 2 a.m. The rebuild moved the scoring prompt to prompts/lead_score.yaml, the model name to an environment variable, and the chain into build_lead_chain(cfg). Now a prompt change is a small YAML diff that the sales-ops lead can review, and the evaluation set of 200 past enquiries runs on every change before deploy.

Follow-up questions to expect

• "How do you let non-engineers edit prompts?" — Store prompts in LangSmith's prompt hub and pull a tagged version at start-up with Client().pull_prompt("lead-score:prod").
• "Why is loads on user input dangerous?" — It creates Python objects from the payload, and a crafted payload can point a model at an attacker's base_url or leak environment secrets.
• "What about saving an agent's state?" — That is conversation state, not the chain: use a LangGraph checkpointer such as PostgresSaver.