examination/topics/qunar-ai-fullstack/ai-structured-output/fill_blank.json

{
  "topic": "ai-structured-output",
  "type": "fill_blank",
  "schema_version": "1.0.0",
  "generated": "2026-09-02T00:00:00Z",
  "questions": [
    {
      "id": "fb-001",
      "type": "fill_blank",
      "difficulty": 2,
      "tags": [
        "ai",
        "structured-output",
        "agent"
      ],
      "question": "JSON Mode 保证模型输出的内容是合法的 ______，但内容本身可能不正确。",
      "explanation": "JSON Mode 强制输出合法 JSON，可靠性中等——格式正确但内容可能有误。",
      "source": null,
      "related": [],
      "answer": [
        "JSON 格式"
      ],
      "answer_rule": "any"
    },
    {
      "id": "fb-002",
      "type": "fill_blank",
      "difficulty": 2,
      "tags": [
        "ai",
        "structured-output",
        "agent"
      ],
      "question": "Structured Output 基于什么技术来保证输出符合 Schema？",
      "explanation": "Structured Output 使用受限解码，在每步 token 生成时屏蔽不合法的 token。",
      "source": null,
      "related": [],
      "answer": [
        "受限解码（Constrained Decoding）"
      ],
      "answer_rule": "any"
    },
    {
      "id": "fb-003",
      "type": "fill_blank",
      "difficulty": 2,
      "tags": [
        "ai",
        "structured-output",
        "agent"
      ],
      "question": "在受限解码过程中，JSON Schema 首先被转换为有限状态自动机（FSA）或______。",
      "explanation": "受限解码的第一步是将 JSON Schema 转换为 FSA 或 CFG 来追踪合法 token 状态。",
      "source": null,
      "related": [],
      "answer": [
        "上下文无关文法（CFG）"
      ],
      "answer_rule": "any"
    },
    {
      "id": "fb-004",
      "type": "fill_blank",
      "difficulty": 2,
      "tags": [
        "ai",
        "structured-output",
        "agent"
      ],
      "question": "Agent 系统的核心循环是：用户输入 → 规划 → 执行 → ______ →（未完成则循环）→ 最终结果。",
      "explanation": "Agent 核心循环：User Input → Planning → Execution → Observation → 循环 → Final Result。",
      "source": null,
      "related": [],
      "answer": [
        "观察"
      ],
      "answer_rule": "any"
    },
    {
      "id": "fb-005",
      "type": "fill_blank",
      "difficulty": 2,
      "tags": [
        "ai",
        "structured-output",
        "agent"
      ],
      "question": "哪种 Agent 模式在执行后进行自我反思，失败时调整策略？",
      "explanation": "Reflexion 模式在执行后自我反思，失败时调整策略，适合需要试错的任务。",
      "source": null,
      "related": [],
      "answer": [
        "Reflexion"
      ],
      "answer_rule": "any"
    },
    {
      "id": "fb-006",
      "type": "fill_blank",
      "difficulty": 2,
      "tags": [
        "ai",
        "structured-output",
        "agent"
      ],
      "question": "ReAct 模式中，Agent 遵循的步骤序列是：思考 → 行动 → 观察 → 思考 → ... → ______。",
      "explanation": "ReAct 交替进行推理和行动，直到得出最终回答（Answer）。",
      "source": null,
      "related": [],
      "answer": [
        "回答"
      ],
      "answer_rule": "any"
    },
    {
      "id": "fb-007",
      "type": "fill_blank",
      "difficulty": 2,
      "tags": [
        "ai",
        "structured-output",
        "agent"
      ],
      "question": "工具链设计中，哪个原则确保相同输入无论调用多少次都产生相同结果？",
      "explanation": "幂等性（Idempotency）意味着相同输入多次调用产生相同结果。",
      "source": null,
      "related": [],
      "answer": [
        "幂等性"
      ],
      "answer_rule": "any"
    },
    {
      "id": "fb-008",
      "type": "fill_blank",
      "difficulty": 2,
      "tags": [
        "ai",
        "structured-output",
        "agent"
      ],
      "question": "对于删除、支付等敏感操作，Agent 工具链推荐使用什么机制？",
      "explanation": "敏感操作应使用人工确认（Human-in-the-Loop）来防止意外的破坏性操作。",
      "source": null,
      "related": [],
      "answer": [
        "Human-in-the-Loop（HITL）/ 人工确认"
      ],
      "answer_rule": "any"
    },
    {
      "id": "fb-009",
      "type": "fill_blank",
      "difficulty": 2,
      "tags": [
        "ai",
        "structured-output",
        "agent"
      ],
      "question": "哪种结构化输出方法灵活性最高但延迟也最高？",
      "explanation": "分步生成（先生成骨架再填充）灵活性最高、可靠性高，但延迟最高。",
      "source": null,
      "related": [],
      "answer": [
        "分步生成"
      ],
      "answer_rule": "any"
    },
    {
      "id": "fb-010",
      "type": "fill_blank",
      "difficulty": 2,
      "tags": [
        "ai",
        "structured-output",
        "agent"
      ],
      "question": "三种低延迟的结构化输出方法是：JSON Mode、Structured Output 和______。",
      "explanation": "JSON Mode、Structured Output 和 Function Calling 都具有低延迟特性。",
      "source": null,
      "related": [],
      "answer": [
        "Function Calling"
      ],
      "answer_rule": "any"
    }
  ]
}