Template Variables and Partials

Where we left off

Lesson 3's template had exactly one blank, {question}. Real templates often need more than one. This lesson adds a second blank, {persona}, and introduces a shortcut for when one blank is fixed for a while but another keeps changing.

Two blanks in one template

prompt = ChatPromptTemplate.from_messages(
    [
        ("system", "You are {persona}. Answer in at most two sentences."),
        ("human", "{question}"),
    ]
)

Nothing conceptually new here beyond Lesson 3, just two placeholders instead of one. Filling them in works the same way, a dictionary with a key matching each blank's name:

filled = prompt.invoke({"persona": "a pirate", "question": "What is LangChain?"})

Both persona and question must be present in the dictionary, or LangChain won't have a value for one of the blanks and will raise an error, same rule as Lesson 3, just now applying to two keys instead of one.

The problem .partial() solves

Imagine persona is decided once, at the start of a session (say, the user picked "a pirate" from a menu), but question is going to be different every single time the user asks something. Passing {"persona": "a pirate", "question": ...} on every single call means repeating persona every time, even though it never actually changes.

pirate_prompt = prompt.partial(persona="a pirate")

.partial() locks in a value for one (or more) of the blanks ahead of time, and hands back a new template that only still has the remaining blank(s) open. pirate_prompt now only needs {"question": ...}, persona is already baked in.

filled_again = pirate_prompt.invoke({"question": "What is a prompt template?"})

Notice this call only passes one key. Run the lesson and you'll see the pirate persona still shapes the answer, even though we never mentioned "pirate" again after building pirate_prompt.

Why this matters beyond saving a few keystrokes

This is the same shape of problem you'll see again in later lessons: some configuration is fixed for a whole session (a persona, a system role, a user's name), while other input changes on every call (the actual question). .partial() is how LangChain lets you separate those two categories cleanly: decide the fixed part once, and only worry about the changing part from then on.