Messages

A conversation is a list. Each item says who spoke and what they said. That is all a message is, and it is the only new idea in this lesson.

Up to now your state has held strings you invented, like ticket and reply. From here it holds a conversation, because that is what you hand a model.

Example

from langchain_core.messages import SystemMessage, HumanMessage, AIMessage

chat = [
    SystemMessage("You are a polite support agent."),
    HumanMessage("I was charged twice"),
    AIMessage("Thanks, I will look into that."),
]

for m in chat:
    print(f"{m.type:<6} {m.content}")

The three you will use most

Type	Who it is	When you write it
`SystemMessage`	The instructions	Once, at the front. It tells the model how to behave.
`HumanMessage`	The person	Every time a user says something.
`AIMessage`	The model	You rarely write these. The model gives them to you.

Every message has a type, which is the short name printed above, and content, which is what was said. There are more fields, and one of them becomes important in lesson 17, but you do not need it yet.

There is a fourth type for the answer a tool gives back. It arrives in lesson 18, once there is a tool to give it.

Why the import says langchain

These come from langchain_core, not from LangGraph. LangGraph is the part that runs your graph. The message types, the models and the tools all come from LangChain, which is the layer underneath. You installed both when you installed LangGraph.

Try it yourself

Add a second HumanMessage and a second AIMessage, so the list reads as a real back and forth.
Print m.type on its own and see the short names.
Try print(chat[1]) and look at everything a message actually carries.

PreviousRouting from inside a node

You understood something today that you didn't yesterday.