Session 4 — LangGraph
Goals
- Move the loop from session 3 into explicit state.
- Distinguish a node that returns a change from code that changes state in place.
- Add a different transition when retrieval does not return the needed source.
- Stop before an action, save the state, and continue after approval.
Why a framework
The loop from session 3 keeps everything in one list. It is hard to stop in the middle, go back to the same step, or test a node without running the whole story. LangGraph gives this a shape: a state object, nodes, and edges.
A node reads the keys it needs and returns only what changed. The framework merges. Changing the list in place breaks the ability to go back to a saved state.
The framework does not decide for you. The stop condition, the budget and the tool permissions stay in your code, exactly as in the direct loop.
A minimal example
The state includes the task, the ids already retrieved, a step counter, and an answer. After the retrieval node, a conditional edge asks:
- Both sources are there: continue to the calculation.
- A source is missing and budget remains: go back to retrieval with a different query.
- The budget ran out: an exit node that returns a stop, not a made-up answer.
The following snippet is a sketch of intent, not code the site installs for you:
def route(state):
if state["answer"]:
return "end"
if state["step"] >= state["budget"]:
return "stop"
if "C:1" not in state["ids"]:
return "retrieve"
return "calculate"And this is how the nodes are connected into a graph. This is a sketch too: the functions retrieve, calculate and stop are yours.
from langgraph.graph import StateGraph, START, END
graph = StateGraph(State)
graph.add_node("retrieve", retrieve)
graph.add_node("calculate", calculate)
graph.add_node("stop", stop)
graph.add_edge(START, "retrieve")
graph.add_conditional_edges("retrieve", route, {
"retrieve": "retrieve", "calculate": "calculate", "stop": "stop", "end": END,
})
graph.add_edge("calculate", END)
graph.add_edge("stop", END)Saving, stopping and resuming
When you compile the graph with a checkpointer, the state is saved after every node, under a conversation id. You can ask the graph to stop before a given node, for example before a node that sends a message. A person reads the saved state and approves, and the run continues from the same point, without repeating the retrievals.
This stop is only useful if the node after it is the only one that performs the action. If an earlier node already sent, the approval comes too late.
What to check in the lab
Run the same question over both sources. Save a snapshot of the state after every node. Turn off the edge that goes back to retrieval, and show that the process stops with one source. Then add a stop before the calculation node, resume the run, and check with the counter that retrieval did not run again.
Check your understanding
Reading
The official documentation: LangGraph, the chapters on persistence and on interrupts. The package version will be published in Moodle before the graded work.
Summary
Found a mistake on this page? Email the lecturer