Message Types

Where we left off

Lesson 1 sent a single bare string:

model.invoke("In one sentence, what is LangChain for?")

That works, but it's limited. There's no way to say "behave a certain way" separately from "here's my actual question", they're smushed together into one plain string. This lesson introduces a better format: a list of labeled messages.

Three message types

from langchain_core.messages import AIMessage, HumanMessage, SystemMessage

• SystemMessage: standing instructions, how the AI should behave. Not something the user said, it's configuration for the whole conversation.
• HumanMessage: something the user (you) said.
• AIMessage: something the AI said back. You saw this already, Lesson 1's response was secretly an AIMessage the whole time, we just never looked at its type directly.

Sending a list instead of a string

messages = [
    SystemMessage("You are a pirate. Answer everything in pirate speak."),
    HumanMessage("What is LangChain?"),
]
response = model.invoke(messages)

.invoke() accepts either a bare string (Lesson 1's way) or a list of messages (this lesson's way). A bare string is actually shorthand, LangChain quietly turns it into a single HumanMessage behind the scenes. Writing the list explicitly is what gives you the ability to add a SystemMessage alongside it, something a bare string can't do on its own.

Run this and the reply comes back in full pirate speak, proof the SystemMessage actually shaped the AI's behavior, separately from the question itself.

A first look at what "memory" will mean

messages.append(response)
messages.append(HumanMessage("Say that again, but in one word."))
second_response = model.invoke(messages)

Here's something worth noticing closely. We took the AI's own reply (response, an AIMessage) and added it back into our list, then added a new question, then sent the whole list again, not just the new question.

The result: the model correctly answers "Arrr!", a one-word pirate summary of what it just said. It only knows what "that" refers to because we resent its own earlier reply back to it, alongside the new question. Nothing was actually remembered by the model itself, this list is the only place that earlier exchange lives, and we chose to include it. Lesson 17 will make this exact mechanism the whole point, but seeing it happen here, early, matters, it demystifies "memory" before it ever gets called by that name.