ai-engineering-from-scratch/phases/14-agent-engineering/13-langgraph-stateful-graphs/quiz.json at main · zhaozhiqiangjianming-ship-it/ai-engineering-from-scratch · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
{
  "lesson": "13-langgraph-stateful-graphs",
  "title": "LangGraph: Stateful Graphs and Durable Execution",
  "questions": [
    {
      "stage": "pre",
      "question": "What does LangGraph treat as the core unit of the agent?",
      "options": [
        "A free-form LLM call",
        "A state machine with typed state, function nodes, and conditional edges",
        "A single tool registry",
        "A vector index"
      ],
      "correct": 1,
      "explanation": "LangGraph models the agent as a state graph: nodes are pure functions, edges are transitions, state is typed and immutable."
    },
    {
      "stage": "pre",
      "question": "Which problem does durable execution solve?",
      "options": [
        "Reducing inference cost",
        "Resuming a 40-step run from step 38 when it fails, with exact state, instead of starting over",
        "Generating embeddings faster",
        "Translating between providers"
      ],
      "correct": 1,
      "explanation": "Checkpoints after every node let the runtime resume from the last successful step."
    },
    {
      "stage": "check",
      "question": "Which of these is NOT one of the three topologies LangGraph supports?",
      "options": [
        "Supervisor",
        "Swarm (peer-to-peer)",
        "Hierarchical (nested subgraphs)",
        "Gradient ring"
      ],
      "correct": 3,
      "explanation": "Topologies are supervisor, swarm, and hierarchical. Gradient ring is not a LangGraph topology."
    },
    {
      "stage": "check",
      "question": "Why must nodes be deterministic for resume to work cleanly?",
      "options": [
        "Providers require determinism",
        "Resume assumes the same inputs produce the same state update; random seeds, wall-clock, and external APIs must be captured",
        "Determinism reduces token cost",
        "It is required by the GIL"
      ],
      "correct": 1,
      "explanation": "If a node depends on uncaptured nondeterminism, resume cannot reconstruct the post-step state."
    },
    {
      "stage": "check",
      "question": "What is a conditional edge?",
      "options": [
        "An edge weighted by training loss",
        "An edge chosen by a function of state, used to branch the graph",
        "An edge that runs only on GPUs",
        "An edge with a TTL"
      ],
      "correct": 1,
      "explanation": "Conditional edges branch based on state; overusing them makes the graph hard to reason about."
    },
    {
      "stage": "post",
      "question": "What goes wrong when checkpoints are too small?",
      "options": [
        "Tool state and memory writes are not recoverable; full state must serialize",
        "The disk fills up",
        "The graph cannot reach END",
        "The model produces shorter answers"
      ],
      "correct": 0,
      "explanation": "Only checkpointing conversation turns leaves tool state and memory writes outside resume's reach."
    },
    {
      "stage": "post",
      "question": "Where does human-in-the-loop fit into LangGraph's design?",
      "options": [
        "It requires a fork of the runtime",
        "Pause before a critical node, surface serialized state to a human, accept modifications, resume; the checkpointer makes this cheap",
        "Only at START and END",
        "Through a separate provider API"
      ],
      "correct": 1,
      "explanation": "Because state is already serialized between nodes, human review and edit is a natural pause-and-resume pattern."
    }
  ]
}