LCEL Demo 4

As LangChain pipelines grow in complexity, you often need to forward data between stages or even modify inputs on the fly. This guide walks you through how to leverage RunnablePassthrough from langchain_core.runnables to:

• Act as a no-op placeholder in a chain.
• Inject or override keys in the input dictionary at runtime with assign.

You’ll build on a simple Prompt → LLM → Parser pipeline and then introduce passthrough steps for flexibility.

1. Basic Prompt → LLM → Parser Chain

Start by creating a minimal chain that sends a prompt to an LLM and parses the raw string output:

from langchain_openai import ChatOpenAI
from langchain_core.prompts import ChatPromptTemplate
from langchain_core.output_parsers import StrOutputParser

# 1. Define a prompt template
prompt = ChatPromptTemplate.from_template(
    "You are a helpful assistant on {topic}.\n"
    "Answer the following question: {question}"
)

# 2. Initialize the LLM client and parser
llm = ChatOpenAI()
output_parser = StrOutputParser()

# 3. Build the chain: Prompt → LLM → Parser
chain = prompt | llm | output_parser

# 4. Invoke with inputs
response = chain.invoke({
    "topic": "movies",
    "question": "Tell me about The Godfather movie"
})
print(response)

Expected output:

"The Godfather" is a classic American crime film released in 1972, directed by Francis Ford Coppola...

2. Adding a No-Op with RunnablePassthrough

You can insert RunnablePassthrough anywhere to forward data unchanged. It’s useful as a placeholder for future logic or debugging:

from langchain_core.runnables import RunnablePassthrough

# Insert at the beginning
chain1 = RunnablePassthrough() | prompt | llm | output_parser
print(chain1.invoke({"topic": "movies", "question": "Tell me about The Godfather movie"}))

# Or at the end
chain2 = prompt | llm | output_parser | RunnablePassthrough()
print(chain2.invoke({"topic": "movies", "question": "Tell me about The Godfather movie"}))

Both invocations return the same result since passthrough does not alter the payload.

3. Verifying the Output

Even with a passthrough at the tail, the result remains identical:

chain_end = prompt | llm | output_parser | RunnablePassthrough()
result = chain_end.invoke({
    "topic": "movies",
    "question": "Tell me about The Godfather movie"
})
print(result)

"The Godfather" is a classic crime film directed by Francis Ford Coppola ...

4. Injecting or Overriding Inputs with assign

The true power of RunnablePassthrough lies in its assign method, which lets you inject or override dictionary keys before they reach downstream components. For instance, you can fix the topic to "movies", so callers only need to supply question:

new_chain = (
    RunnablePassthrough() |
    RunnablePassthrough.assign(topic=lambda _: "movies") |
    prompt |
    llm |
    output_parser
)

# Now the chain accepts only 'question'
response = new_chain.invoke({"question": "Tell me about Inception"})
print(response)

Under the hood:

1. The first passthrough forwards {"question": ...}.
2. assign(topic=lambda _: "movies") adds "topic": "movies".
3. The prompt, LLM, and parser receive both keys.

You can isolate and test assign by itself:

test_chain = (
    RunnablePassthrough() |
    RunnablePassthrough.assign(topic=lambda _: "movies")
)
result = test_chain.invoke({"question": "Tell me about Inception"})
print(result)

Expected output:

{'question': 'Tell me about Inception', 'topic': 'movies'}

Common assign Patterns

5. Summary

You’ve learned how to:

• Use RunnablePassthrough as a no-op placeholder for debugging or future pipeline steps.
• Employ assign to inject or override keys in your chain inputs.
• Maintain a clean and testable pipeline structure with minimal boilerplate.

You can even wrap arbitrary Python functions using RunnableLambda for custom transformations.

Links and References

• LangChain Documentation
• ChatOpenAI on GitHub
• RunnablePassthrough Source
• LangChain Output Parsers