AWS AIP-C01 开源练习 · 高难 · 第 02 套
英文原题与中文机器翻译。本批暂不提供日文题文；日语界面显示英文原文。
开源练习，非 AWS 官方试题。按作者参考答案判分，解析附官方文档链接。
Long documents mix several topics. Each chunk must hold ONE coherent topic for precise matching; there is no need to also pass a larger parent context. Which strategy is BEST?
长文档混合了多个主题。每个块必须包含一个连贯的主题以进行精确匹配；无需传递更大的父上下文。哪种策略最好？
- A
Semantic chunking.
语义分块。
- B
Hierarchical (parent-child) chunking.
分层（父子）分块。
- C
Fixed-size chunking.
固定大小的分块。
- D
Only increasing chunk overlap.
仅增加块重叠。
解析 / Explanation
正确答案 / Answer: A
A. Semantic chunking splits on meaning so each chunk is one coherent topic, which fixes topic bleed without an extra parent layer.
B. Hierarchical adds a parent-context layer that is unnecessary here and does not specifically fix topic boundaries.
C. Fixed sizes cut across topic boundaries, the root problem.
D. Overlap helps boundary continuity but does not make each chunk a single coherent topic.
A. 语义分块根据含义进行分割，因此每个块都是一个连贯的主题，这无需额外的父层即可修复主题出血。
B. 分层添加了一个父上下文层，这在此处是不必要的，并且不会专门修复主题边界。
C. 固定大小跨越了主题边界，这是根本问题。
D. 重叠有助于边界连续性，但不会使每个块成为一个连贯的主题。
Answers are frequently cut off because relevant sentences fall right at the boundary between two chunks. Which adjustment most directly fixes this?
答案经常被切断，因为相关句子恰好位于两个块之间的边界。哪种调整最能直接解决这个问题？
- A
Add chunk overlap so adjacent chunks share boundary text.
添加块重叠，以便相邻块共享边界文本。
- B
Reduce the chunk size.
减小块大小。
- C
Switch to hierarchical chunking.
切换到分层分块。
- D
Increase the number of results returned.
增加返回结果的数量。
解析 / Explanation
正确答案 / Answer: A
A. Overlap repeats boundary text so information spanning a boundary is captured in at least one chunk.
B. Smaller chunks create MORE boundaries and can make boundary loss worse.
C. Hierarchical chunking is heavier than needed and does not specifically target boundary cut-offs.
D. Returning more chunks does not fix content split across a boundary.
A. 重叠重复边界文本，因此跨越边界的信息至少被捕获在一个块中。
B. 较小的块会创建更多的边界，并使边界损失更严重。
C. 分层分块比需要的更重，并且不专门针对边界截止。
D. 返回更多块并不能解决跨边界分割的内容。
After upgrading to a newer embedding model, retrieval quality collapsed even though documents are present. What must the team do?
升级到更新的嵌入模型后，即使存在文档，检索质量也会下降。团队必须做什么？
- A
Re-embed (re-index) the entire corpus with the new model so document and query vectors share the same space.
使用新模型重新嵌入（重新索引）整个语料库，以便文档和查询向量共享相同的空间。
- B
Re-embed only the queries with the new model.
仅使用新模型重新嵌入查询。
- C
Clear the application cache.
清除应用程序缓存。
- D
Lower the generation temperature.
降低生成温度。
解析 / Explanation
正确答案 / Answer: A
A. Query and document embeddings must come from the same model; switching models requires re-embedding the whole corpus.
B. Mismatched query and document spaces still produce poor retrieval.
C. Caching is unrelated to embedding-space mismatch.
D. Temperature does not affect embedding compatibility.
A. 查询和文档嵌入必须来自同一模型；切换模型需要重新嵌入整个语料库。
B. 不匹配的查询和文档空间仍然会导致检索效果不佳。
C. 缓存与嵌入空间不匹配无关。
D. 温度不会影响嵌入兼容性。
A growing vector index is too expensive. The team wants to cut storage and query cost while keeping acceptable recall. Which change helps MOST?
不断增长的向量索引太昂贵了。该团队希望降低存储和查询成本，同时保持可接受的召回率。哪项改变最有帮助？
- A
Use an embedding model with smaller output dimensions (for example, Titan Text Embeddings V2) and pick a smaller dimension.
使用输出尺寸较小的嵌入模型（例如 Titan Text Embeddings V2）并选择较小的尺寸。
- B
Use the maximum embedding dimension available.
使用可用的最大嵌入尺寸。
- C
Store each chunk twice for redundancy.
将每个块存储两次以实现冗余。
- D
Switch to a multimodal embedding model.
切换到多模态嵌入模型。
解析 / Explanation
正确答案 / Answer: A
A. Smaller-dimension embeddings shrink the index and query cost while keeping acceptable recall.
B. Maximum dimensions increase storage and cost.
C. Duplication increases cost without improving accuracy.
D. Multimodal embeddings are for image+text retrieval and do not reduce text storage cost.
A. 较小维度的嵌入可以缩小索引和查询成本，同时保持可接受的召回率。
B. 最大尺寸增加了存储和成本。
C. 重复会增加成本，但不会提高准确性。
D. 多模态嵌入用于图像+文本检索，不会降低文本存储成本。
Evaluation shows the correct passage EXISTS in the store but is sometimes NOT returned, so the model lacks the facts to answer. Which change should the team try FIRST?
评估显示商店中存在正确的通道，但有时不会返回，因此模型缺乏要回答的事实。团队应该首先尝试哪个改变？
- A
Increase the number of results (top-k) returned per query.
增加每个查询返回的结果数 (top-k)。
- B
Add a reranker model.
添加重新排序模型。
- C
Reduce the chunk overlap.
减少块重叠。
- D
Lower the temperature.
降温。
解析 / Explanation
正确答案 / Answer: A
A. Returning more candidates raises recall so the correct passage is more likely included.
B. A reranker reorders retrieved results but cannot rank a passage that was never retrieved.
C. Less overlap does not improve whether the missing passage is retrieved.
D. Temperature affects generation, not retrieval recall.
A. 返回更多候选者可以提高召回率，因此更有可能包含正确的段落。
B. 重新排序器对检索到的结果进行重新排序，但无法对从未检索到的段落进行排序。
C. 较少的重叠并不能改善丢失的段落是否被检索。
D. 温度影响生成，而不影响检索回忆。
Chunks were configured very large, so retrieval returns passages where the relevant sentence is buried among unrelated text and match quality is weak. Which change is BEST?
块配置得非常大，因此检索返回的段落中相关句子被隐藏在不相关的文本中，并且匹配质量很弱。哪种改变最好？
- A
Reduce the chunk size to more focused units of meaning.
将块大小缩小为更集中的意义单元。
- B
Increase the chunk size further.
进一步增加块大小。
- C
Add more chunk overlap.
添加更多块重叠。
- D
Increase the embedding dimensions.
增加嵌入尺寸。
解析 / Explanation
正确答案 / Answer: A
A. Smaller, focused chunks produce embeddings that better represent specific content, improving match precision.
B. Larger chunks dilute the embedding more and worsen precision.
C. Overlap addresses boundaries, not the dilution caused by oversized chunks.
D. Dimension size does not fix oversized, diluted chunks.
A. 较小的、集中的块产生的嵌入可以更好地表示特定内容，从而提高匹配精度。
B. 较大的块会进一步稀释嵌入并降低精度。
C. 重叠解决的是边界问题，而不是超大块造成的稀释问题。
D. 尺寸大小不能修复过大、稀释的块。
A knowledge base ingests many short, self-contained FAQ entries, where each file is already one complete question-and-answer. Which chunking choice is BEST?
知识库包含许多简短的、独立的常见问题解答条目，其中每个文件已经是一个完整的问答。哪种分块选择最好？
- A
Use a no-chunking strategy so each short file is one chunk.
使用无分块策略，使每个短文件都是一个块。
- B
Apply small fixed-size chunking to every file.
对每个文件应用小的固定大小分块。
- C
Merge all FAQ files into one chunk.
将所有常见问题解答文件合并为一个块。
- D
Use hierarchical chunking.
使用分层分块。
解析 / Explanation
正确答案 / Answer: A
A. When each file is already atomic, no chunking preserves the complete meaning per entry.
B. Splitting atomic FAQ entries fragments coherent answers.
C. Merging unrelated entries dilutes each entry's embedding.
D. Hierarchical structure is unnecessary for already-atomic entries.
A. 当每个文件已经是原子的时，任何分块都无法保留每个条目的完整含义。
B. 拆分原子常见问题解答条目会导致连贯的答案碎片化。
C. 合并不相关的条目会淡化每个条目的嵌入。
D. 对于已经原子的条目来说，层次结构是不必要的。
A catalog must be searchable by EITHER a text query or an image query against the same items. Which embedding approach supports this?
目录必须可通过针对相同项目的文本查询或图像查询进行搜索。哪种嵌入方法支持这一点？
- A
Use a multimodal embedding model (such as Amazon Titan Multimodal Embeddings) so images and text share one vector space.
使用多模态嵌入模型（例如 Amazon Titan Multimodal Embeddings），以便图像和文本共享一个向量空间。
- B
Use a text-only embedding model for everything.
对所有内容都使用纯文本嵌入模型。
- C
Index Amazon Rekognition labels only.
仅索引 Amazon Rekognition 标签。
- D
Embed base64 image strings as text.
将 Base64 图像字符串嵌入为文本。
解析 / Explanation
正确答案 / Answer: A
A. Multimodal embeddings place images and text in a shared space, enabling cross-modal similarity search.
B. A text-only model cannot embed images for similarity search.
C. Label detection is not a shared semantic embedding space for text-and-image retrieval.
D. Embedding base64 as text does not capture visual semantics.
A. 多模态嵌入将图像和文本放置在共享空间中，从而实现跨模态相似性搜索。
B. 纯文本模型无法嵌入图像进行相似性搜索。
C. 标签检测不是用于文本和图像检索的共享语义嵌入空间。
D. 将 base64 嵌入为文本不会捕获视觉语义。
A team wants to raise the overall quality of what their RAG system retrieves and answers. Which TWO techniques are most effective? (Select TWO.)
团队希望提高 RAG 系统检索和应答内容的整体质量。哪两种技术最有效？ （选择两项。）
- A
Choose a chunking strategy (semantic or hierarchical) aligned to the documents' structure.
选择与文档结构一致的分块策略（语义或分层）。
- B
Add a reranking step to reorder retrieved candidates by relevance.
添加重新排序步骤以按相关性对检索到的候选者重新排序。
- C
Raise the generation temperature for variety.
提高品种的生成温度。
- D
Increase maxTokens on the response.
增加响应的 maxTokens。
- E
Always return only the single top result.
始终仅返回单个顶部结果。
解析 / Explanation
正确答案 / Answer: A,B
A. Structure-aligned chunking produces more coherent, relevant chunks and better retrieval.
B. A reranker moves the most relevant passages to the top for generation.
C. Higher temperature increases randomness and does not improve retrieval quality.
D. Output length does not change which passages are retrieved.
E. Returning just one result can drop relevant passages and reduce answer quality.
A. 结构对齐的分块可以产生更连贯、相关的分块和更好的检索。
B. 重新排序器将最相关的段落移动到顶部以进行生成。
C. 较高的温度会增加随机性，但不会提高检索质量。
D. 输出长度不会改变检索的段落。
E. 仅返回一个结果可能会删除相关段落并降低答案质量。
A team must A/B test two wordings of a production prompt and revert instantly if one underperforms, WITHOUT redeploying application code. Which option is BEST?
团队必须对生产提示的两个措辞进行 A/B 测试，如果其中一个措辞表现不佳，则立即恢复，而无需重新部署应用程序代码。哪个选项最好？
- A
Amazon Bedrock Prompt Management with versioned prompts and variants.
具有版本化提示和变体的 Amazon Bedrock 提示管理。
- B
Hard-code both prompts behind an application feature flag.
将两个提示硬编码在应用程序功能标志后面。
- C
Deploy two Lambda functions, each hard-coding one prompt.
部署两个 Lambda 函数，每个函数硬编码一个提示。
- D
Adjust the temperature to switch behavior.
调节温度以切换行为。
解析 / Explanation
正确答案 / Answer: A
A. Managed prompt versions/variants allow comparison and instant switch/rollback with no code deploy.
B. A feature flag still requires a code deploy to ship the prompts and lacks managed versioning/evaluation.
C. This is custom, requires code deploys, and has no managed prompt versioning.
D. Temperature does not select between prompt wordings.
A. 托管提示版本/变体允许比较和即时切换/回滚，无需部署代码。
B. 功能标志仍然需要代码部署来发送提示，并且缺乏托管版本控制/评估。
C. 这是自定义的，需要部署代码，并且没有托管提示版本控制。
D. 温度不选择提示文字。
A developer must cap response length to control cost and avoid overly long chat outputs. Which inference setting is correct?
开发人员必须限制响应长度以控制成本并避免聊天输出过长。哪种推理设置是正确的？
- A
Set the maxTokens (maximum output tokens) parameter.
设置 maxTokens（最大输出令牌）参数。
- B
Raise the temperature.
提高温度。
- C
Increase top-k.
增加top-k。
- D
Add unrelated stop sequences.
添加不相关的停止序列。
解析 / Explanation
正确答案 / Answer: A
A. maxTokens directly bounds output length and cost.
B. Temperature changes randomness, not maximum length.
C. Top-k changes token diversity, not length.
D. Unrelated stop sequences do not reliably cap general length.
A. maxTokens 直接限制输出长度和成本。
B. 温度变化是随机的，而不是最大长度。
C. Top-k 改变了 token 的多样性，而不是长度。
D. 不相关的停止序列不能可靠地限制总体长度。
Durable behavior rules (role, tone, safety constraints) must govern EVERY turn of a conversation, separate from each user question. Where should they go?
持久的行为规则（角色、语气、安全约束）必须控制对话的每一轮，与每个用户问题分开。他们应该去哪里？
- A
In the system prompt.
系统提示中。
- B
Appended by the user to each message.
由用户附加到每条消息中。
- C
In the temperature setting.
在温度设定中。
- D
In the stop sequences list.
在停止序列列表中。
解析 / Explanation
正确答案 / Answer: A
A. The system prompt sets persistent instructions that apply across the conversation, separate from per-turn user messages.
B. Relying on users to repeat rules is inconsistent and unreliable.
C. Temperature is a sampling parameter, not a place for rules.
D. Stop sequences end generation; they do not encode behavior rules.
A. 系统提示设置适用于整个对话的持久指令，与每轮用户消息分开。
B. 依赖用户重复规则是不一致且不可靠的。
C. 温度是一个采样参数，而不是规则的地方。
D. 停止序列结束生成；他们不编码行为规则。
Outputs must follow a precise structured format, instructions alone have not worked, and the team needs a fix QUICKLY without a training step. Which technique is BEST?
输出必须遵循精确的结构化格式，仅靠指令是行不通的，团队需要在没有培训步骤的情况下快速修复。哪种技术最好？
- A
Add a few worked examples (few-shot exemplars) of the exact format to the prompt.
在提示中添加一些精确格式的工作示例（少数示例）。
- B
Fine-tune the model on formatting examples.
在格式化示例上微调模型。
- C
Increase the temperature.
提高温度。
- D
Shorten the prompt to one sentence.
将提示缩短为一句话。
解析 / Explanation
正确答案 / Answer: A
A. Few-shot examples show the model the precise structure to imitate and require no training.
B. Fine-tuning is heavier and slower than needed when few-shot prompting can fix formatting quickly.
C. Higher temperature increases variability and undermines format consistency.
D. Over-shortening removes the structure cues the model needs.
A. 很少的例子向模型展示了可以模仿的精确结构，并且不需要训练。
B. 当几次提示可以快速修复格式时，微调会比需要的更重且更慢。
C. 较高的温度会增加变异性并破坏格式的一致性。
D. 过度缩短会消除模型所需的结构线索。
Several applications reuse the same prompt with only a few values changing per request, and the team wants central management with runtime value injection. Which approach is BEST?
多个应用程序重复使用相同的提示，每个请求仅更改几个值，并且团队希望通过运行时值注入进行集中管理。哪种方法最好？
- A
Amazon Bedrock Prompt Management templates with input variables.
带有输入变量的 Amazon Bedrock 提示管理模板。
- B
Copy the prompt into each application and edit by hand.
将提示复制到每个应用程序中并手动编辑。
- C
Concatenate strings ad hoc in each service.
在每个服务中临时连接字符串。
- D
Raise the temperature to make the prompt flexible.
提高温度，使提示灵活。
解析 / Explanation
正确答案 / Answer: A
A. Parameterized managed templates are reused centrally and take values at invocation time.
B. Duplicated prompts drift and are hard to govern centrally.
C. Ad hoc concatenation has no central governance or versioning.
D. Temperature does not parameterize or manage prompts.
A. 参数化托管模板可集中重用，并在调用时获取值。
B. 重复的提示会漂移并且难以集中管理。
C. 即席串联没有中央治理或版本控制。
D. 温度不会参数化或管理提示。
A developer needs a MANAGED, low-code way to build and version a multi-step generative workflow with branching (classify, then retrieve, then generate). Which Amazon Bedrock feature is BEST?
开发人员需要一种托管的低代码方式来构建具有分支的多步骤生成工作流并对其进行版本控制（分类，然后检索，然后生成）。哪个 Amazon Bedrock 功能最好？
- A
Amazon Bedrock Flows (prompt flows).
亚马逊基岩流（即时流）。
- B
Amazon Bedrock Agents.
亚马逊基岩代理。
- C
Hand-built AWS Step Functions wiring each step.
手工构建的 AWS Step Functions 连接每个步骤。
- D
A single InvokeModel call.
单个 InvokeModel 调用。
解析 / Explanation
正确答案 / Answer: A
A. Bedrock Flows visually build and version multi-step, branching workflows chaining prompts, knowledge bases, and other components with little code.
B. Agents provide autonomous tool-using reasoning, which is heavier and different from a defined, branching prompt-chain workflow.
C. Step Functions can orchestrate but require more custom build than a managed flow builder.
D. One call cannot orchestrate a multi-step, branching workflow.
A. Bedrock Flows 以可视化方式构建和版本化多步骤、分支工作流程，链接提示、知识库和其他组件，只需很少的代码。
B. 代理提供自主的使用工具的推理，这比定义的分支提示链工作流程更重且不同。
C. Step Functions 可以进行编排，但比托管流程构建器需要更多的自定义构建。
D. 一次调用无法编排多步骤、分支的工作流程。
A regression suite needs the same prompt to produce as close to identical output as possible on each run. Which SINGLE parameter change most increases determinism?
回归套件需要相同的提示，以便在每次运行时产生尽可能接近相同的输出。哪个单一参数的变化最能增加确定性？
- A
Set the temperature to 0.
将温度设置为 0。
- B
Set a very high top-k.
设置非常高的top-k。
- C
Increase maxTokens.
增加最大令牌数。
- D
Raise the temperature to 1.
将温度升至 1。
解析 / Explanation
正确答案 / Answer: A
A. A temperature of 0 makes decoding effectively greedy, maximizing run-to-run consistency.
B. A high top-k widens sampling and increases variability.
C. Output length does not control determinism.
D. A high temperature maximizes randomness.
A. 温度为 0 会使解码有效地贪婪，从而最大限度地提高运行间的一致性。
B. 高 top-k 扩大了采样范围并增加了变异性。
C. 输出长度不控制确定性。
D. 高温使随机性最大化。
In a RAG assistant, the team wants to reduce fabricated answers by making the model rely only on the supplied context and admit when the answer is absent. Which prompt technique helps MOST?
在 RAG 助手中，团队希望通过使模型仅依赖于提供的上下文来减少捏造的答案，并在答案不存在时承认。哪种提示技术最有帮助？
- A
Instruct the model to answer only from the provided context and to say it does not know when the context lacks the answer.
指示模型仅根据提供的上下文进行回答，并表示它不知道上下文何时缺少答案。
- B
Raise the temperature for more creative answers.
提高温度以获得更有创意的答案。
- C
Remove the retrieved context to shorten prompts.
删除检索到的上下文以缩短提示。
- D
Instruct the model to always produce an answer.
指示模型始终产生答案。
解析 / Explanation
正确答案 / Answer: A
A. Explicit grounding plus an allowed abstention reduces fabrication when the context does not contain the answer.
B. More creativity increases fabrication.
C. Without context the model must guess, increasing hallucination.
D. Forcing an answer encourages fabrication when information is missing.
A. 当上下文不包含答案时，明确的基础加上允许的弃权可以减少捏造行为。
B. 更多的创造力会增加制造能力。
C. 在没有上下文的情况下，模型必须猜测，从而增加幻觉。
D. 当信息缺失时，强行给出答案会鼓励捏造事实。
A developer must prevent unbounded, runaway generations that drive up cost and ramble past the answer. Which TWO controls directly help? (Select TWO.)
开发人员必须防止无限的、失控的一代，这会增加成本并超出答案。哪两个控件有直接帮助？ （选择两项。）
- A
Set a sensible maxTokens limit.
设置合理的 maxTokens 限制。
- B
Define stop sequences that end generation at a known marker.
定义在已知标记处结束生成的停止序列。
- C
Increase the temperature.
提高温度。
- D
Increase top-k substantially.
大幅增加top-k。
- E
Enlarge the context window.
放大上下文窗口。
解析 / Explanation
正确答案 / Answer: A,B
A. maxTokens caps how many tokens can be generated, bounding cost and runaway length.
B. Stop sequences halt generation at a defined boundary so the model does not ramble on.
C. Higher temperature does not limit length and can make output more verbose.
D. Top-k affects diversity, not length.
E. The context window governs input size, not output length.
A. maxTokens 限制了可以生成的代币数量、边界成本和失控长度。
B. 停止序列在定义的边界处停止生成，这样模型就不会继续下去。
C. 较高的温度不会限制长度，并且会使输出更加冗长。
D. Top-k 影响多样性，而不是长度。
E. 上下文窗口控制输入大小，而不是输出长度。
A model's outputs occasionally include bizarre, very low-probability words. The team wants to remove that long tail of unlikely tokens while keeping SOME variety. Which single adjustment is BEST?
模型的输出有时会包含奇怪的、概率极低的单词。该团队希望消除不太可能的代币的长尾，同时保持一定的多样性。哪种单一调整最好？
- A
Lower top-p (nucleus sampling) so only the most probable tokens are eligible.
降低 top-p（核心采样），因此只有最可能的标记才符合资格。
- B
Set temperature to 0.
将温度设置为 0。
- C
Increase top-k substantially.
大幅增加top-k。
- D
Add stop sequences.
添加停止序列。
解析 / Explanation
正确答案 / Answer: A
A. A lower top-p trims the low-probability tail while still allowing variety among the top tokens, directly removing bizarre word choices.
B. Temperature 0 removes the tail but also eliminates the variety the team wants to keep.
C. A larger top-k widens the candidate pool, making rare tokens more likely, not less.
D. Stop sequences end generation; they do not filter the token probability distribution.
A. 较低的 top-p 会修剪低概率的尾部，同时仍然允许顶部标记之间存在多样性，从而直接消除奇怪的单词选择。
B. 温度 0 会去除尾部，但也会消除团队想要保留的多样性。
C. 较大的 top-k 扩大了候选池，使稀有代币出现的可能性更大，而不是更少。
D. 停止序列结束生成；它们不过滤令牌概率分布。
An Amazon Bedrock agent must call an internal order-management API to look up and update orders. What is the standard way to give it that capability?
Amazon Bedrock 代理必须调用内部订单管理 API 来查找和更新订单。赋予它这种能力的标准方法是什么？
- A
Define an action group backed by a Lambda function, described with an API or function schema.
定义由 Lambda 函数支持的操作组，并使用 API 或函数架构进行描述。
- B
Put the API documentation in the agent's instructions.
将API文档放在代理的说明中。
- C
Associate a knowledge base with the agent.
将知识库与代理相关联。
- D
Raise the agent model's temperature.
提高代理模型的温度。
解析 / Explanation
正确答案 / Answer: A
A. Action groups declare callable operations via a schema and execute them through a Lambda when the agent decides to act.
B. Instructions describe behavior but provide no executable tool to call the API.
C. A knowledge base grounds answers in documents; it does not execute API actions.
D. Temperature has nothing to do with tool or API access.
A. 操作组通过模式声明可调用操作，并在代理决定执行操作时通过 Lambda 执行它们。
B. 指令描述了行为，但不提供调用 API 的可执行工具。
C. 知识库以文档中的答案为基础；它不执行 API 操作。
D. 温度与工具或 API 访问无关。
For compliance, the company's OWN application must execute tool calls, while the agent still decides which action to take. Which capability supports this?
为了合规性，公司自己的应用程序必须执行工具调用，而代理仍然决定采取哪个操作。哪种能力支持这个？
- A
Configure the action group to return control to the caller, which executes the action and returns the result.
配置操作组以将控制权返回给调用者，调用者执行操作并返回结果。
- B
Use an action group with a Lambda the agent invokes directly.
将操作组与代理直接调用的 Lambda 结合使用。
- C
Disable the agent's reasoning.
禁用代理的推理。
- D
Associate a knowledge base.
关联知识库。
解析 / Explanation
正确答案 / Answer: A
A. Return of control hands the chosen action and parameters back to the calling application to execute, rather than the agent invoking a Lambda.
B. With a Lambda action group the agent executes the action, which is exactly what the compliance rule forbids.
C. That would remove the agent's ability to decide which action to take.
D. A knowledge base is for retrieval, not for delegating execution to the caller.
A. 控制权返回将所选操作和参数交还给调用应用程序来执行，而不是调用 Lambda 的代理。
B. 通过 Lambda 操作组，代理执行操作，这正是合规性规则所禁止的。
C. 这将消除代理决定采取哪种行动的能力。
D. 知识库用于检索，而不是用于将执行委托给调用者。
An agent must ground its answers in company policy documents, with the least custom code. What should the developer do?
代理必须将答案建立在公司政策文件中，并使用最少的自定义代码。开发商应该做什么？
- A
Associate an Amazon Bedrock knowledge base with the agent.
将 Amazon Bedrock 知识库与代理关联。
- B
Add an action group to fetch documents from S3.
添加一个操作组以从 S3 获取文档。
- C
Paste the entire policy library into the agent instructions.
将整个策略库粘贴到代理指令中。
- D
Fine-tune the agent's model weekly on the policies.
每周根据政策微调代理模型。
解析 / Explanation
正确答案 / Answer: A
A. Associating a knowledge base lets the agent retrieve and ground responses without custom retrieval code.
B. Action groups execute operations; grounding from a corpus is what a knowledge base provides.
C. Instructions cannot scale to a full library or stay current.
D. Repeated fine-tuning is heavy and unnecessary versus a knowledge base.
A. 关联知识库使代理可以检索并接地响应，而无需自定义检索代码。
B. 行动组执行操作；知识库提供的是语料库的基础。
C. 说明无法扩展到完整的库或保持最新状态。
D. 与知识库相比，重复的微调是繁重且不必要的。
Each user's agent must remember preferences and prior context ACROSS separate sessions, as a managed AgentCore capability. Which component is correct?
每个用户的代理必须记住跨单独会话的首选项和先前上下文，作为托管 AgentCore 功能。哪个组件是正确的？
- A
AgentCore Memory.
代理核心内存。
- B
Per-invocation session state only.
仅每次调用会话状态。
- C
AgentCore Gateway.
AgentCore 网关。
- D
A custom DynamoDB table you build and maintain.
您构建和维护的自定义 DynamoDB 表。
解析 / Explanation
正确答案 / Answer: A
A. AgentCore Memory provides managed short-term and long-term memory that persists across sessions.
B. Session state lasts within a single session; it does not persist across sessions.
C. Gateway exposes APIs as tools; it is not a memory store.
D. That works but is not the managed AgentCore capability the scenario asks for.
A. AgentCore Memory 提供跨会话持续存在的托管短期和长期内存。
B. 会话状态在单个会话内持续；它不会跨会话持续存在。
C. 网关将 API 作为工具公开；它不是一个内存存储。
D. 这可行，但不是场景所要求的托管 AgentCore 功能。
A company wants to expose its MANY existing internal REST APIs and Lambda functions to agents as secure, governed tools without writing per-API integration code. Which AgentCore capability fits?
一家公司希望将其许多现有的内部 REST API 和 Lambda 函数作为安全、受管控的工具向代理公开，而无需编写每个 API 集成代码。哪种 AgentCore 功能适合？
- A
AgentCore Gateway.
AgentCore 网关。
- B
Defining a separate action group for each API in every agent.
为每个代理中的每个 API 定义单独的操作组。
- C
AgentCore Code Interpreter.
AgentCore 代码解释器。
- D
AgentCore Identity.
代理核心身份。
解析 / Explanation
正确答案 / Answer: A
A. Gateway turns existing APIs/Lambdas into agent-ready tools (for example via MCP) with built-in authentication and governance.
B. Per-agent action groups require wiring each API repeatedly; Gateway centralizes existing APIs as managed tools.
C. Code Interpreter executes code; it is not an API-to-tool exposure layer.
D. Identity manages access/credentials, not exposure of APIs as tools.
A. Gateway 将现有 API/Lambda 转变为具有内置身份验证和治理的代理就绪工具（例如通过 MCP）。
B. 每个代理操作组需要重复连接每个 API； Gateway 将现有 API 作为托管工具进行集中。
C. 代码解释器执行代码；它不是 API 到工具的暴露层。
D. 身份管理访问/凭证，而不是将 API 作为工具公开。
An analytics agent must safely execute Python to perform calculations and data transformations, isolated from other workloads, as a managed capability. Which AgentCore component is correct?
分析代理必须安全地执行 Python 来执行计算和数据转换，作为托管功能与其他工作负载隔离。哪个 AgentCore 组件是正确的？
- A
AgentCore Code Interpreter.
AgentCore 代码解释器。
- B
AgentCore Gateway.
AgentCore 网关。
- C
A self-built Lambda action group for code execution.
自建的 Lambda 操作组，用于执行代码。
- D
AgentCore Browser.
AgentCore 浏览器。
解析 / Explanation
正确答案 / Answer: A
A. Code Interpreter gives the agent a managed, isolated sandbox to run code.
B. Gateway exposes external tools; it is not a code-execution sandbox.
C. You would have to build and secure the sandbox yourself; Code Interpreter is the managed option.
D. Browser automates web pages, not arbitrary code execution.
A. 代码解释器为代理提供了一个托管的、隔离的沙箱来运行代码。
B. 网关暴露外部工具；它不是代码执行沙箱。
C. 您必须自己构建并保护沙箱；代码解释器是托管选项。
D. 浏览器自动化网页，而不是任意代码执行。
An agent must NAVIGATE websites and interact with web pages to complete tasks, in a managed, secure environment. Which AgentCore capability fits?
代理必须在受管理的安全环境中导航网站并与网页交互才能完成任务。哪种 AgentCore 功能适合？
- A
AgentCore Browser.
AgentCore 浏览器。
- B
AgentCore Gateway.
AgentCore 网关。
- C
AgentCore Code Interpreter.
AgentCore 代码解释器。
- D
AgentCore Memory.
代理核心内存。
解析 / Explanation
正确答案 / Answer: A
A. The Browser tool provides a managed, secure cloud browser the agent uses to navigate and interact with pages.
B. Gateway exposes APIs as tools; it does not provide interactive web browsing.
C. Code Interpreter runs code; it is not a browser.
D. Memory persists context, not web interaction.
A. 浏览器工具提供了一个托管的安全云浏览器，代理使用它来导航页面并与页面交互。
B. 网关将 API 作为工具公开；它不提供交互式网页浏览。
C. 代码解释器运行代码；它不是浏览器。
D. 内存保留的是上下文，而不是网络交互。
Agents must securely access third-party SaaS services ON BEHALF OF users, handling tokens and consent, without a custom credential broker. Which AgentCore capability addresses this?
代理必须代表用户安全地访问第三方 SaaS 服务，处理令牌和同意，而无需自定义凭证代理。哪种 AgentCore 功能可以解决这个问题？
- A
AgentCore Identity.
代理核心身份。
- B
AWS Secrets Manager.
AWS 秘密管理器。
- C
Amazon Cognito user pools alone.
仅 Amazon Cognito 用户池。
- D
AgentCore Gateway.
AgentCore 网关。
解析 / Explanation
正确答案 / Answer: A
A. AgentCore Identity manages secure identity and access for agents, including authenticating to and accessing third-party services on a user's behalf.
B. Secrets Manager stores secrets but is not the agent identity/consent broker for acting on a user's behalf.
C. Cognito authenticates end users to your app; it is not the agent's outbound identity to third-party services.
D. Gateway exposes tools; it is not the identity/consent layer.
A. AgentCore Identity 管理代理的安全身份和访问，包括代表用户验证和访问第三方服务。
B. Secrets Manager 存储机密，但不是代表用户行事的代理身份/同意代理。
C. Cognito 对您的应用程序的最终用户进行身份验证；它不是代理向第三方服务提供的出站身份。
D. 网关公开工具；它不是身份/同意层。
A research agent has long-running, multi-minute sessions and must be deployed serverlessly with per-session isolation, regardless of the framework it was built in. Which option is BEST?
研究代理具有长时间运行、多分钟的会话，并且必须以每会话隔离的方式进行无服务器部署，无论其构建于何种框架中。哪个选项是最好的？
- A
Amazon Bedrock AgentCore Runtime.
Amazon Bedrock AgentCore 运行时。
- B
AWS Lambda.
AWS Lambda。
- C
A self-managed Amazon ECS on Fargate service.
Fargate 服务上的自我管理 Amazon ECS。
- D
A SageMaker real-time endpoint.
SageMaker 实时端点。
解析 / Explanation
正确答案 / Answer: A
A. AgentCore Runtime hosts agents from any framework serverlessly with session isolation and support for long-running, streaming workloads.
B. Lambda's execution time limit and lack of per-session agent isolation make it unsuitable for long, stateful agent sessions.
C. ECS works but requires building and operating the serving infrastructure.
D. SageMaker endpoints serve models, not framework-agnostic agent runtimes.
A. AgentCore Runtime 无服务器地托管来自任何框架的代理，具有会话隔离并支持长时间运行的流式工作负载。
B. Lambda 的执行时间限制和缺乏每会话代理隔离使其不适合长时间、有状态的代理会话。
C. ECS 可以工作，但需要构建和运营服务基础设施。
D. SageMaker 端点服务于模型，而不是与框架无关的代理运行时。
An operations team must trace WHY an agent chose a tool and inspect each reasoning step, tool call, and latency in production. Which AgentCore capability provides this?
运营团队必须追踪代理选择工具的原因，并检查每个推理步骤、工具调用和生产中的延迟。哪种 AgentCore 功能可以提供此功能？
- A
AgentCore Observability.
AgentCore 可观察性。
- B
AWS CloudTrail.
AWS CloudTrail。
- C
Bedrock model invocation logging.
基岩模型调用日志记录。
- D
AgentCore Memory.
代理核心内存。
解析 / Explanation
正确答案 / Answer: A
A. Observability captures detailed traces of the agent's reasoning steps and tool invocations for debugging and monitoring.
B. CloudTrail records API calls, not the agent's internal reasoning steps and tool decisions.
C. Invocation logging captures model prompts/responses, not the agent's step-by-step tool orchestration.
D. Memory persists context; it is not a tracing tool.
A. 可观察性捕获代理推理步骤和工具调用的详细跟踪，以进行调试和监控。
B. CloudTrail 记录 API 调用，而不是代理的内部推理步骤和工具决策。
C. 调用日志记录捕获模型提示/响应，而不是代理的逐步工具编排。
D. 记忆保留上下文；它不是一个跟踪工具。
A workflow needs ONE coordinating agent to dynamically delegate subtasks to several specialized agents (billing, shipping, returns) and combine results. Which approach is BEST?
工作流程需要一个协调代理将子任务动态委托给多个专门代理（计费、运输、退货）并合并结果。哪种方法最好？
- A
Use multi-agent collaboration with a supervisor agent coordinating specialized sub-agents.
使用多代理协作，由主管代理协调专门的子代理。
- B
Build the workflow in Amazon Bedrock Flows.
在 Amazon Bedrock Flows 中构建工作流程。
- C
Put every capability in one agent's prompt.
将每一项功能都纳入一个代理的提示中。
- D
Run independent agents with no coordination.
运行独立代理，无需协调。
解析 / Explanation
正确答案 / Answer: A
A. Multi-agent collaboration lets a supervisor route subtasks to specialized agents and aggregate their outputs dynamically.
B. Flows is a fixed, low-code orchestration of steps; it is not dynamic agent-to-agent delegation and reasoning.
C. A single monolithic agent provides no structured delegation or specialization.
D. Without a supervisor there is no way to combine sub-agent results.
A. 多代理协作让主管可以将子任务分配给专门代理并动态聚合其输出。
B. Flows 是一种固定的、低代码的步骤编排；它不是动态代理到代理的委派和推理。
C. 单个整体代理不提供结构化委托或专业化。
D. 如果没有主管，就无法合并子代理结果。
An agent can issue refunds, and a human must approve before any refund executes. Which Bedrock agent feature enforces this directly?
代理可以发放退款，并且在执行任何退款之前必须经过人工批准。哪个基岩代理功能直接强制执行此操作？
- A
Enable user confirmation (require confirmation) on the sensitive action.
对敏感操作启用用户确认（需要确认）。
- B
Use return of control for the action.
使用操作的控制权返回。
- C
Add a note in the system prompt to confirm first.
在系统提示中添加备注，先确认。
- D
Set temperature to 0.
将温度设置为 0。
解析 / Explanation
正确答案 / Answer: A
A. User confirmation pauses the agent for explicit human approval before the action runs.
B. Return of control changes WHERE execution happens (the caller), but it is not itself a built-in human-approval gate.
C. Prompt instructions do not enforce a hard approval gate.
D. Temperature does not create an approval step.
A. 用户确认会在操作运行之前暂停代理以进行明确的人工批准。
B. 控制权的返回会改变执行发生的位置（调用者），但它本身并不是内置的人工批准门。
C. 提示指令不会强制执行硬性批准门。
D. 温度不会产生批准步骤。
A developer wants an OPEN-SOURCE, code-first, model-driven framework to build tool-using agents in a few lines of Python and then deploy them to AgentCore Runtime. Which option fits?
开发人员需要一个开源、代码优先、模型驱动的框架，用几行 Python 代码构建使用工具的代理，然后将它们部署到 AgentCore Runtime。哪个选项适合？
- A
The Strands Agents SDK.
Strands Agent SDK。
- B
Amazon Bedrock Agents configured in the console.
在控制台中配置的 Amazon Bedrock 代理。
- C
Amazon Q Developer.
亚马逊 Q 开发人员。
- D
AWS Glue.
AWS 胶水。
解析 / Explanation
正确答案 / Answer: A
A. Strands Agents is AWS's open-source, model-driven SDK for building agents in code and deploying them to AgentCore Runtime.
B. Bedrock Agents is a managed, configuration-driven service, not the open-source code-first SDK described.
C. Q Developer is a coding assistant, not an agent-building runtime SDK.
D. Glue is an ETL service, unrelated to building agents.
A. Strands Agents 是 AWS 的开源模型驱动 SDK，用于在代码中构建代理并将其部署到 AgentCore Runtime。
B. Bedrock Agents 是一项托管、配置驱动的服务，而不是所述的开源代码优先 SDK。
C. Q Developer 是一个编码助手，而不是一个代理构建运行时 SDK。
D. Glue 是一项 ETL 服务，与构建代理无关。
An agent's responses must be filtered for harmful content and PII regardless of which tools or knowledge bases it uses. Which approach is BEST?
无论代理使用哪种工具或知识库，都必须过滤其响应中的有害内容和 PII。哪种方法最好？
- A
Attach an Amazon Bedrock Guardrail to the agent.
将 Amazon Bedrock Guardrail 连接到代理。
- B
Write the safety rules into the agent's instructions.
将安全规则写入代理商的说明中。
- C
Filter content inside each action group Lambda.
过滤每个操作组 Lambda 内的内容。
- D
Lower the agent's temperature.
降低试剂的温度。
解析 / Explanation
正确答案 / Answer: A
A. A guardrail attached to the agent applies content filtering and PII handling consistently across the agent's interactions.
B. Instructions are not a reliable, enforced safety control and can be bypassed.
C. Per-tool custom filtering is inconsistent and misses responses that do not go through that Lambda.
D. Temperature does not filter harmful content or PII.
A. 连接到代理的护栏在代理的交互中一致地应用内容过滤和 PII 处理。
B. 指令不是可靠的、强制的安全控制，并且可以被绕过。
C. 每个工具的自定义过滤不一致，并且会错过不经过该 Lambda 的响应。
D. 温度不会过滤有害内容或 PII。
An application needs synchronous, low-latency predictions for SMALL payloads with steady, predictable traffic. Which SageMaker AI option is BEST?
应用程序需要对具有稳定、可预测流量的小型有效负载进行同步、低延迟预测。哪种 SageMaker AI 选项最好？
- A
A real-time inference endpoint.
实时推理端点。
- B
Serverless Inference.
无服务器推理。
- C
Asynchronous Inference.
异步推理。
- D
Batch transform.
批量变换。
解析 / Explanation
正确答案 / Answer: A
A. Real-time endpoints give persistent, low-latency synchronous inference, ideal for steady small-payload traffic.
B. Serverless can cold-start and does not support GPUs, adding latency variability not wanted for steady low-latency traffic.
C. Async adds queueing latency meant for long-running or large requests, not small fast ones.
D. Batch transform is offline; it cannot serve synchronous requests.
A. 实时端点提供持久、低延迟的同步推理，非常适合稳定的小负载流量。
B. 无服务器可以冷启动并且不支持 GPU，从而增加了稳定的低延迟流量所不需要的延迟可变性。
C. 异步增加了排队延迟，适用于长时间运行或大型请求，而不是小型快速请求。
D. 批量变换离线；它无法服务同步请求。
A CPU model serves intermittent traffic with long idle gaps, can tolerate occasional cold starts, and must not incur idle cost. Which option is MOST cost-effective?
CPU 模型提供具有较长空闲间隙的间歇性流量，可以容忍偶尔的冷启动，并且不得产生空闲成本。哪种选择最具成本效益？
- A
SageMaker Serverless Inference.
SageMaker 无服务器推理。
- B
An always-on real-time endpoint.
始终在线的实时端点。
- C
Asynchronous Inference sized for peak.
异步推理的大小适合峰值。
- D
Provisioned, always-on GPU instance.
已配置、始终在线的 GPU 实例。
解析 / Explanation
正确答案 / Answer: A
A. Serverless scales (including to zero) for intermittent CPU workloads, billing only while serving.
B. An always-on endpoint pays for the long idle periods.
C. Provisioning for peak wastes capacity for intermittent traffic.
D. A continuously running GPU is the most expensive and unnecessary for a CPU workload.
A. 无服务器可针对间歇性 CPU 工作负载进行扩展（包括扩展到零），仅在服务时计费。
B. 始终在线的端点需要为长时间空闲期付出代价。
C. 为间歇性流量提供高峰浪费容量。
D. 对于 CPU 工作负载来说，连续运行的 GPU 是最昂贵且不必要的。
Requests carry payloads up to ~1 GB and take several minutes on a GPU; responses can be delivered asynchronously. Which SageMaker option is BEST?
请求携带高达约 1 GB 的有效负载，并且在 GPU 上需要几分钟时间；响应可以异步传递。哪种 SageMaker 选项最好？
- A
Asynchronous Inference.
异步推理。
- B
Real-time Inference.
实时推理。
- C
Serverless Inference.
无服务器推理。
- D
Batch transform.
批量变换。
解析 / Explanation
正确答案 / Answer: A
A. Async supports large payloads (up to ~1 GB), long processing, and GPU instances with queued, asynchronous responses.
B. Real-time endpoints cap payloads near 25 MB and target short synchronous calls, so large/long requests fail.
C. Serverless does not support GPUs and has small payload limits.
D. Batch transform is for bulk offline scoring, not on-demand per-request handling.
A. 异步支持大负载（高达约 1 GB）、长时间处理以及具有排队异步响应的 GPU 实例。
B. 实时端点将有效负载限制为接近 25 MB，并以短同步调用为目标，因此大/长请求会失败。
C. Serverless 不支持 GPU，并且负载限制较小。
D. 批量转换用于批量离线评分，而不是按请求按需处理。
A team deploys several large language models to GPU instances and wants to pack multiple models/copies onto shared accelerators, scale each independently, and scale unused ones to ZERO. Which capability fits?
一个团队将多个大型语言模型部署到 GPU 实例，并希望将多个模型/副本打包到共享加速器上，独立扩展每个模型，并将未使用的模型扩展为零。适合哪种能力？
- A
SageMaker inference components.
SageMaker 推理组件。
- B
A multi-model endpoint (MME).
多模型端点 (MME)。
- C
One real-time endpoint per model.
每个模型一个实时端点。
- D
Serverless Inference.
无服务器推理。
解析 / Explanation
正确答案 / Answer: A
A. Inference components place multiple models/copies on shared instances, allocate resources per component, scale independently, and can scale to zero.
B. MME loads many (typically smaller) models on demand behind one endpoint but is not designed for GPU-packing LLMs with independent per-model scaling and scale-to-zero.
C. Separate endpoints cannot pack models onto shared accelerators with independent scaling.
D. Serverless does not support the GPUs these LLMs need.
A. 推理组件将多个模型/副本放置在共享实例上，为每个组件分配资源，独立扩展，并且可以扩展到零。
B. MME 在一个端点后面按需加载许多（通常较小）模型，但不是为具有独立的每个模型缩放和缩放至零的 GPU 封装 LLM 设计的。
C. 单独的端点无法将模型打包到具有独立扩展的共享加速器上。
D. 无服务器不支持这些法学硕士所需的 GPU。
A company must host HUNDREDS of small, similar models (one per customer) behind as few endpoints as possible, loading each on demand to save cost. Which option is BEST?
公司必须在尽可能少的端点后面托管数百个小型相似模型（每个客户一个），按需加载每个模型以节省成本。哪个选项最好？
- A
A multi-model endpoint (MME).
多模型端点 (MME)。
- B
SageMaker inference components.
SageMaker 推理组件。
- C
One dedicated endpoint per model.
每个模型一个专用端点。
- D
Batch transform per model.
每个模型的批量转换。
解析 / Explanation
正确答案 / Answer: A
A. MME hosts many models behind one endpoint, loading them on demand and sharing resources to cut cost.
B. Inference components target packing and independently scaling (often larger) models, not the classic many-small-models-on-demand pattern.
C. Hundreds of endpoints multiply cost and management.
D. Batch transform is offline scoring, not on-demand hosting.
A. MME 在一个端点后面托管许多模型，按需加载它们并共享资源以降低成本。
B. 推理组件的目标是打包和独立缩放（通常更大）模型，而不是经典的按需多个小模型模式。
C. 数百个端点会增加成本和管理成本。
D. 批量转换是离线评分，不是按需托管。
Real-time endpoint traffic swings widely between day and night, and the team wants instances added and removed automatically based on load. What should they configure?
实时端点流量在白天和晚上之间波动很大，团队希望根据负载自动添加和删除实例。他们应该配置什么？
- A
Automatic scaling (for example, target tracking on invocations per instance).
自动扩展（例如，每个实例的调用目标跟踪）。
- B
A fixed instance count sized for peak.
针对峰值大小的固定实例计数。
- C
Switch to batch transform during the day.
白天切换到批量转换。
- D
Use a larger instance type only.
仅使用较大的实例类型。
解析 / Explanation
正确答案 / Answer: A
A. Endpoint auto scaling adjusts instance count against a target metric, matching capacity to demand.
B. A fixed peak size wastes money at night and cannot adapt automatically.
C. Batch transform is offline and cannot serve real-time daytime traffic.
D. A bigger instance does not automatically scale with demand.
A. 端点自动扩展根据目标指标调整实例计数，使容量与需求相匹配。
B. 固定的峰值大小在晚上会浪费金钱，而且无法自动适应。
C. 批量转换是离线的，无法服务于白天的实时流量。
D. 更大的实例不会根据需求自动扩展。
Before promoting a new model version, the team wants to send a COPY of live production traffic to it and compare results WITHOUT affecting any users. Which capability is correct?
在推广新模型版本之前，团队希望向其发送实时生产流量的副本并比较结果而不影响任何用户。哪种能力是正确的？
- A
Shadow testing (shadow variants).
影子测试（影子变体）。
- B
A canary deployment with deployment guardrails.
带有部署护栏的金丝雀部署。
- C
A blue/green switch all at once.
蓝色/绿色同时切换。
- D
Batch transform on yesterday's data.
对昨天的数据进行批量转换。
解析 / Explanation
正确答案 / Answer: A
A. Shadow testing mirrors live traffic to a shadow variant so you validate against real requests with no impact on user-facing responses.
B. Canary routes a slice of REAL user traffic to the new version, so some users are affected, unlike shadow testing.
C. An all-at-once cutover exposes all users immediately and is not a no-impact comparison.
D. Offline scoring does not exercise the new version against live traffic.
A. 影子测试将实时流量镜像到影子变体，以便您根据真实请求进行验证，而不会影响面向用户的响应。
B. 金丝雀将一部分真实用户流量路由到新版本，因此与影子测试不同，一些用户会受到影响。
C. 一次性全部切换会立即暴露所有用户，并且不是无影响的比较。
D. 离线评分不会针对实时流量使用新版本。
When updating a production endpoint, the team wants to shift a small percentage of REAL traffic first and automatically roll back if errors spike. Which capability is correct?
更新生产端点时，团队希望首先转移一小部分真实流量，并在错误激增时自动回滚。哪种能力是正确的？
- A
Deployment guardrails with a canary (or linear) traffic-shifting strategy and automatic rollback.
具有金丝雀（或线性）流量转移策略和自动回滚的部署护栏。
- B
Shadow testing.
阴影测试。
- C
Recreate the endpoint in place.
就地重新创建端点。
- D
Increase the instance count during the switch.
增加切换期间的实例计数。
解析 / Explanation
正确答案 / Answer: A
A. Deployment guardrails gradually shift live traffic, watch alarms, and roll back automatically if the new version misbehaves.
B. Shadow testing mirrors traffic without serving users, so it does not perform a controlled production rollout.
C. An in-place replace has no gradual shift or automatic rollback and risks downtime.
D. More instances does not provide controlled traffic shifting or rollback.
A. 部署护栏会逐渐转移实时流量、监视警报，并在新版本出现问题时自动回滚。
B. 影子测试镜像流量而不服务用户，因此它不执行受控的生产部署。
C. 就地替换不会逐渐转变或自动回滚，并且存在停机风险。
D. 更多实例不提供受控的流量转移或回滚。
A developer wants to deploy a popular open foundation model fast, with a tested container and example notebooks. Which option is BEST?
开发人员希望通过经过测试的容器和示例笔记本快速部署流行的开放基础模型。哪个选项最好？
- A
Deploy the model from Amazon SageMaker JumpStart.
从 Amazon SageMaker JumpStart 部署模型。
- B
Build a custom serving container from scratch.
从头开始构建自定义服务容器。
- C
Use batch transform to expose an endpoint.
使用批量转换来公开端点。
- D
Use AWS Glue to host the model.
使用 AWS Glue 托管模型。
解析 / Explanation
正确答案 / Answer: A
A. JumpStart provides prebuilt, deployable models with curated containers and notebooks for fast deployment.
B. Building everything yourself is the slowest path.
C. Batch transform does not create an online endpoint.
D. Glue is an ETL service and cannot host a model endpoint.
A. JumpStart 提供预构建的可部署模型以及精选的容器和笔记本，以实现快速部署。
B. 自己构建一切是最慢的路径。
C. 批量转换不会创建在线端点。
D. Glue 是一项 ETL 服务，无法托管模型端点。
A fixed dataset already in Amazon S3 must be scored in a SINGLE offline run with no persistent endpoint, at the lowest cost. Which option is BEST?
Amazon S3 中已有的固定数据集必须在没有持久端点的单次离线运行中以最低成本进行评分。哪个选项最好？
- A
A batch transform job.
批量转换作业。
- B
A real-time endpoint invoked in a loop.
循环调用的实时端点。
- C
Asynchronous Inference per record.
每条记录的异步推理。
- D
Serverless Inference.
无服务器推理。
解析 / Explanation
正确答案 / Answer: A
A. Batch transform scores a whole S3 dataset in one offline run and releases compute afterward, with no endpoint to pay for.
B. Looping over a persistent endpoint is more expensive and complex than one batch job.
C. Per-record async adds overhead versus a single batch over the dataset.
D. Serverless targets intermittent online requests, not bulk offline scoring.
A. 批量转换在一次离线运行中对整个 S3 数据集进行评分，然后释放计算，无需支付任何端点费用。
B. 循环持久端点比一个批处理作业更昂贵、更复杂。
C. 与数据集上的单个批次相比，每条记录的异步会增加开销。
D. 无服务器的目标是间歇性在线请求，而不是批量离线评分。
Profiling a model served with DJL on an 8-GPU instance shows the model fits entirely within 4 GPUs and per-instance concurrency is low. Which change BEST improves GPU utilization?
对 8-GPU 实例上使用 DJL 提供的模型进行分析表明，该模型完全适合 4 个 GPU，并且每个实例的并发性较低。哪项更改最能提高 GPU 利用率？
- A
Use a tensor parallelism degree of 4 so two model replicas run on each 8-GPU instance.
使用张量并行度 4，以便在每个 8-GPU 实例上运行两个模型副本。
- B
Use a tensor parallelism degree of 8.
使用张量并行度 8。
- C
Add more instances and spread requests.
添加更多实例并传播请求。
- D
Enable speculative decoding.
启用推测解码。
解析 / Explanation
正确答案 / Answer: A
A. Since the model fits in 4 GPUs, a tensor-parallel degree of 4 places two replicas per instance, using all 8 GPUs.
B. Degree 8 serves a single replica across all GPUs, leaving half the capacity idle since the model fits in 4.
C. More instances does not raise per-instance utilization or concurrency.
D. Speculative decoding reduces latency, not GPU utilization.
A. 由于该模型适合 4 个 GPU，因此张量并行度为 4 时，每个实例都会放置两个副本，并使用全部 8 个 GPU。
B. 8 度在所有 GPU 上提供单个副本，由于该模型适合 4 度，因此留下一半的容量闲置。
C. 更多实例不会提高每个实例的利用率或并发性。
D. 推测性解码会减少延迟，但不会减少 GPU 利用率。
Logs show the real maximum sequence length is about 10x smaller than the configured value and concurrency per GPU is low on a DJL-served model. Which change BEST increases throughput per GPU?
日志显示实际最大序列长度比配置值小大约 10 倍，并且在 DJL 服务的模型上每个 GPU 的并发性较低。哪项更改最能提高每个 GPU 的吞吐量？
- A
Reduce the configured maximum sequence length to allow a larger rolling batch size.
减少配置的最大序列长度以允许更大的滚动批量大小。
- B
Increase the maximum sequence length.
增加最大序列长度。
- C
Lower the rolling batch size.
降低轧制批量大小。
- D
Enable speculative decoding.
启用推测解码。
解析 / Explanation
正确答案 / Answer: A
A. Lowering the max sequence length frees memory so more requests fit in each rolling batch, increasing throughput and concurrency.
B. A larger max sequence length reserves more memory per request and reduces batch size.
C. A smaller batch reduces throughput.
D. Speculative decoding targets latency, not batch-size-driven throughput.
A. 降低最大序列长度可以释放内存，因此每个滚动批次可以容纳更多请求，从而提高吞吐量和并发性。
B. 较大的最大序列长度为每个请求保留更多内存并减少批量大小。
C. 较小的批次会降低吞吐量。
D. 推测性解码的目标是延迟，而不是批量大小驱动的吞吐量。
An enterprise must upgrade many applications from an old Java version to a current LTS release, automating the repetitive code changes at scale. Which Amazon Q Developer capability is BEST?
企业必须将许多应用程序从旧的 Java 版本升级到当前的 LTS 版本，从而大规模地自动执行重复的代码更改。哪种 Amazon Q 开发人员功能最好？
- A
Amazon Q Developer code transformation.
Amazon Q Developer 代码转换。
- B
Amazon Q Developer inline code suggestions.
Amazon Q 开发人员内联代码建议。
- C
Manually editing each project.
手动编辑每个项目。
- D
Amazon Macie.
亚马逊麦西.
解析 / Explanation
正确答案 / Answer: A
A. Code transformation automates large-scale upgrades such as moving codebases to a newer Java version.
B. Inline suggestions speed typing in the editor but do not perform an automated, repository-wide upgrade.
C. Manual edits across many applications are slow and error-prone, which automation avoids.
D. Macie classifies sensitive data and is unrelated to code upgrades.
A. 代码转换可自动进行大规模升级，例如将代码库移动到较新的 Java 版本。
B. 内联建议可加快编辑器中的输入速度，但不会执行自动的存储库范围内的升级。
C. 许多应用程序中的手动编辑速度缓慢且容易出错，而自动化可以避免这种情况。
D. Macie 对敏感数据进行分类，与代码升级无关。
A team wants an assistant that takes a natural-language feature request, plans it, and implements the change across MULTIPLE files. Which Amazon Q Developer capability fits?
团队需要一个助手来接受自然语言功能请求，对其进行规划，并跨多个文件实现更改。哪种 Amazon Q Developer 功能适合？
- A
The Amazon Q Developer feature-development agent (/dev).
Amazon Q Developer 功能开发代理 (/dev)。
- B
Single-line inline completion.
单行内联完成。
- C
Amazon QuickSight Q.
亚马逊 QuickSight Q。
- D
Amazon Comprehend.
亚马逊理解。
解析 / Explanation
正确答案 / Answer: A
A. The /dev agent analyzes the codebase, proposes a plan, and applies coordinated multi-file changes.
B. Inline completion helps at the cursor but cannot plan and implement a multi-file feature.
C. QuickSight Q answers questions over BI data, not multi-file code changes.
D. Comprehend is an NLP service, not a coding agent.
A. /dev 代理分析代码库，提出计划，并应用协调的多文件更改。
B. 内联补全有助于光标，但无法规划和实现多文件功能。
C. QuickSight Q 回答有关 BI 数据的问题，而不是有关多文件代码更改的问题。
D. Comprehend 是一种 NLP 服务，而不是编码代理。
Developers want their assistant to scan SOURCE CODE for vulnerabilities and suggest fixes inside the IDE. Which option is correct?
开发人员希望他们的助手扫描源代码中的漏洞并在 IDE 内提出修复建议。哪个选项是正确的？
- A
Amazon Q Developer security scanning.
Amazon Q 开发人员安全扫描。
- B
Amazon Inspector.
亚马逊检查员。
- C
Amazon GuardDuty.
亚马逊 GuardDuty。
- D
Amazon Macie.
亚马逊麦西.
解析 / Explanation
正确答案 / Answer: A
A. Q Developer scans code for vulnerabilities and suggests remediations in the developer workflow.
B. Inspector assesses workloads and software for vulnerabilities operationally, not in-IDE source-code scanning for developers.
C. GuardDuty detects account/workload threats, not source-code vulnerabilities.
D. Macie classifies sensitive data, not code vulnerabilities.
A. Q Developer 扫描代码中的漏洞并在开发人员工作流程中提出修复建议。
B. Inspector 在操作上评估工作负载和软件是否存在漏洞，而不是为开发人员进行 IDE 源代码扫描。
C. GuardDuty 检测帐户/工作负载威胁，而不是源代码漏洞。
D. Macie 对敏感数据进行分类，而不是对代码漏洞进行分类。
To raise coverage quickly, a team wants to automatically GENERATE unit tests for existing functions. Which option is BEST?
为了快速提高覆盖率，团队希望自动生成现有功能的单元测试。哪个选项最好？
- A
Amazon Q Developer unit test generation.
Amazon Q Developer 单元测试生成。
- B
AWS CodeBuild.
AWS 代码构建。
- C
Single-line inline completion only.
仅单行内联完成。
- D
Amazon QuickSight.
亚马逊 QuickSight。
解析 / Explanation
正确答案 / Answer: A
A. Q Developer generates unit tests for existing code with minimal manual effort.
B. CodeBuild runs builds/tests but does not author tests.
C. Inline completion does not systematically generate a test suite for existing functions.
D. QuickSight is for dashboards, not test generation.
A. Q Developer 以最少的手动工作为现有代码生成单元测试。
B. CodeBuild 运行构建/测试，但不编写测试。
C. 内联完成不会系统地为现有功能生成测试套件。
D. QuickSight 适用于仪表板，而不是测试生成。
A team wants automated test generation and quality checks to run CONTINUOUSLY as code is merged, not ad hoc. Which approach is BEST?
团队希望自动化测试生成和质量检查在代码合并时持续运行，而不是临时运行。哪种方法最好？
- A
Integrate Amazon Q Developer automated test generation into the CI/CD pipeline.
将 Amazon Q Developer 自动测试生成集成到 CI/CD 管道中。
- B
Run Amazon Q Developer only manually in the IDE when remembered.
仅当记住时才在 IDE 中手动运行 Amazon Q Developer。
- C
Hold periodic manual review cycles instead.
相反，保持定期的手动审核周期。
- D
Add Amazon Inspector to the pipeline for unit tests.
将 Amazon Inspector 添加到管道以进行单元测试。
解析 / Explanation
正确答案 / Answer: A
A. Embedding Q Developer test generation in CI/CD continuously validates changes and catches defects early.
B. Ad hoc IDE use is not continuous and misses merges.
C. Manual cycles are slow and not continuous.
D. Inspector finds vulnerabilities; it does not generate unit tests.
A. 在 CI/CD 中嵌入 Q Developer 测试生成可以持续验证更改并及早发现缺陷。
B. Ad hoc IDE 使用不连续并且会错过合并。
C. 手动循环缓慢且不连续。
D. 检查员发现漏洞；它不生成单元测试。