How do you implement unit tests for LangChain chains?

What you need to know

The testing layers

A unit test with fakes

1from langchain_core.documents import Document2from langchain_core.language_models.fake_chat_models import FakeListChatModel3from langchain_core.runnables import RunnableLambda45def test_rag_puts_retrieved_text_in_prompt():6    seen = {}7    def spy(prompt_value):                 # records the prompt sent to the model8        seen["prompt"] = prompt_value.to_string()9        return prompt_value1011    llm = FakeListChatModel(responses=["You get 30 days [refunds.md]"])12    retriever = RunnableLambda(lambda q: [13        Document(page_content="Refunds within 30 days.", metadata={"source": "refunds.md"})])1415    chain = build_rag_chain(retriever, RunnableLambda(spy) | llm)16    out = chain.invoke("What is the refund window?")1718    assert "Refunds within 30 days." in seen["prompt"]19    assert "[refunds.md]" in out["answer"]20    assert out["context"][0].metadata["source"] == "refunds.md"

• The fake model returns its responses in order, with no network.
• The spy step checks what the model would have received — the most important thing in a RAG test.
• Assertions are about structure (source present, context passed), not wording.

What else to unit test

• Parsers and validation — feed malformed text or objects; assert a clean error or repair.
• Tools — call the tool function directly with valid and invalid arguments (tool.invoke({...})); mock the external API.
• Agents — use a fake chat model that returns an AIMessage with tool_calls, then assert the right tool ran with the right arguments and that its result reached the next model turn.
• Security — the retriever for tenant A never returns tenant B's documents.
• Error paths — make a fake raise TimeoutError and assert the fallback message.

Quality in CI

With pip install "langsmith[pytest]", tests marked @pytest.mark.langsmith log inputs, outputs and feedback to LangSmith, so an evaluation suite can run as ordinary pytest with results tracked over time.

A real-life example

A finance assistant has a get_stock_price tool and an answer chain. Its tests used to call the real model and the real stock API: 4 minutes per run, flaky because prices changed, and about ₹800 a day in API costs across the team.

The rewrite: unit tests with FakeListChatModel, a fake tool-calling model for the agent, and a stubbed price API that returns fixed prices. The suite of 120 tests runs in 5 seconds. One test caught a real bug — when the stock API timed out, the agent used to say "The price is unavailable" and invent a number; the error-path test now asserts no digits appear in that reply. A nightly job runs 40 real-model evaluation cases and posts scores to the team channel.

Follow-up questions to expect

• "How do you fake a model that calls tools?" — GenericFakeChatModel with a list of AIMessages that include tool_calls; with create_agent you may need a small subclass whose bind_tools returns itself.
• "How do you test prompts?" — Render them with sample inputs and snapshot the result; any change is then visible in review.
• "Can you assert on LLM output at all?" — On real models, use checks that tolerate wording: schema validity, required facts present, LLM-judge scores above a threshold.