Answers with a fixed shape

A TypedDict describes your state but never checks it. If you would rather be told when something is wrong, use a Pydantic model instead.

Everything you have written so far used TypedDict, which is a description and nothing more. Put a number where a string belongs and it will sail through and break somewhere later. A Pydantic model is the same description with the checking turned on.

from pydantic import BaseModel
from langgraph.graph import StateGraph, START

class State(BaseModel):
    ticket: str
    urgent: bool = False

def triage(state: State):
    return {"urgent": "charge" in state.ticket}

builder = StateGraph(State)
builder.add_node("triage", triage)
builder.add_edge(START, "triage")

print(builder.compile().invoke({"ticket": "charged twice"}))

Two differences from a TypedDict. The node reads state.ticket with a dot instead of square brackets, and urgent has a default, so you did not have to pass it in.

What you get for it

from pydantic import BaseModel, ValidationError
from langgraph.graph import StateGraph, START

class State(BaseModel):
    ticket: str

def triage(state: State):
    return {"ticket": state.ticket.upper()}

builder = StateGraph(State)
builder.add_node("triage", triage)
builder.add_edge(START, "triage")

try:
    builder.compile().invoke({"ticket": 12345})
except ValidationError as e:
    print("rejected at the door:", e.errors()[0]["msg"])

The run stopped before any node ran. With a TypedDict that same call would have reached triage and failed on .upper(), several steps away from the thing that was actually wrong.

Which to use

The other kind of fixed shape

There is a related idea worth knowing the name of. You can also ask the model to answer in a fixed shape rather than in prose, by handing with_structured_output a schema.

from pydantic import BaseModel

class Triage(BaseModel):
    category: str
    urgent: bool

sorted_ticket = model.with_structured_output(Triage).invoke(messages)
print(sorted_ticket.category, sorted_ticket.urgent)

That one has no output on this page, because our stand-in model cannot do it. It works by binding the schema as a tool, so it needs a model that really understands tools. Come back to it in lesson 30, once you have a real model.