Wrong value assigned to the "question" at chain pipeline

[DmitryRoss]

knowledge-graph-rag/kgraph_rag/roman_emp_graph_rag.py

Line 233 in 0fbd7e2

"question": RunnablePassthrough(),

First of all, thank you for such good course, I've learned really a lot from you. Respect!

I found a bug at "question": RunnablePassthrough(),
Here you are assigning entire input dictionary to the "question". On the other hand "question" is used at prompt Question: {question}

I've built a small debug function to output what LLM takes as a prompt:

def debug_prompt(prompt_template, name="PROMPT DEBUG"):
    """
    Creates runnable that prints formatted prompt before sending it to the LLM

    :param prompt_template: Your ChatPromptTemplate or PromptTemplate.
    :param name: Label for output.
    """
    def printer(x):
        print(f"\n=== {name} ===")
        formatted_prompt = prompt_template.format(**x)
        print(formatted_prompt)
        print("="*30)
        return x  # returns dict w/o changes to not break down the pipeline
    
    return RunnableLambda(printer)

Added it to the chain pipeline:

chain = (
    RunnableParallel(
        {
            "context": _search_query | retriever,
            "question": RunnablePassthrough(), # replace to fix "question": RunnableLambda(lambda x: x["question"]), 
        }
    )
    | debug_prompt(prompt)
    | prompt
    | chat
    | StrOutputParser()
)

Here is how Question look:
Question: {'question': 'What did Romulus?', 'chat_history': [('Who was the first emperor?', 'Augustus was the first emperor.')]}
Use natural language and be concise.
Answer:

If I were you, I'de replaced "question": RunnablePassthrough(), with "question": RunnableLambda(lambda x: x["question"]). That worked fine for me.

Again, thank you for your work it made me smarter :)

[onestardao]

Thanks for surfacing this — what you’re hitting is exactly what we’ve tracked internally as Issue #11: Chain Pipeline Param Binding Drift.

The core problem is that when question is defined at one stage, but bound downstream using a mismatched key or raw dict, the pipeline passes the wrong payload into the prompt. This can lead to responses that look unrelated or inconsistent.

We’ve catalogued this case already. If it’d help, I can share our checklist entry that shows the exact conditions where the binding drifts and how to patch it. Would you like me to send that over?