AWS AIP-C01 开源练习 · 高难 · 第 01 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。
A production Bedrock workload has steady, high, predictable demand for 12 business hours and near-zero traffic overnight. The team must guarantee dedicated capacity during business hours (no throttling) at the lowest total cost, keeping the same model and API. Which approach is BEST?
生产基岩工作负载在 12 个工作小时内具有稳定、高且可预测的需求，并且夜间流量接近于零。团队必须以最低的总成本保证工作时间内的专用容量（无节流），并保持相同的模型和 API。哪种方法最好？
- A
Use Provisioned Throughput without a long-term commitment, creating model units for business hours and deleting them overnight.
使用预置吞吐量而无需长期承诺，为工作时间创建模型单元并在一夜之间将其删除。
- B
Use Provisioned Throughput with a 6-month commitment sized for peak.
使用预置吞吐量和 6 个月的峰值承诺。
- C
Use on-demand inference for all traffic.
对所有流量使用按需推理。
- D
Use cross-Region inference to add capacity during business hours.
使用跨区域推理在工作时间内增加容量。
解析 / Explanation
正确答案 / Answer: A
A. No-commitment Provisioned Throughput guarantees dedicated capacity while active and is billed only while provisioned, so releasing units overnight avoids paying for the idle period.
B. A committed term is cheapest per hour but you pay for the reserved units overnight too, so with near-zero overnight traffic it is not the lowest total cost.
C. On-demand has no idle cost but cannot guarantee dedicated capacity during peak hours, so it fails the no-throttling requirement.
D. Cross-Region inference spreads on-demand traffic for throughput/availability but does not reserve guaranteed capacity, and it is still on-demand priced.
A. 无承诺预置吞吐量可保证活动时的专用容量，并且仅在预置时计费，因此隔夜释放单位可以避免为空闲期付费。
B. 承诺期限每小时最便宜，但您也需要为预留单位隔夜付费，因此，在夜间流量接近于零的情况下，这并不是最低的总成本。
C. 按需模式没有闲置成本，但无法保证高峰时段的专用容量，无法满足不限流的要求。
D. 跨区域推理会按需分配流量以提高吞吐量/可用性，但不会预留保证容量，并且仍按按需定价。
A synchronous, customer-facing Bedrock feature must become more resilient to transient errors and to a single-Region impairment, per the Reliability pillar. Which TWO measures are MOST effective? (Select TWO.)
根据可靠性支柱，面向客户的同步基岩功能必须对瞬态错误和单区域损伤具有更强的弹性。哪两项措施最有效？ （选择两项。）
- A
Implement retries with exponential backoff and jitter for throttling and transient 5xx errors.
通过指数退避和抖动实现重试，以抑制和瞬态 5xx 错误。
- B
Use cross-Region inference (or multi-Region failover) so requests can be served if one Region is impaired.
使用跨区域推理（或多区域故障转移），以便在一个区域受损时可以为请求提供服务。
- C
Place a single-Region Amazon SQS queue in front of the model call.
将单区域 Amazon SQS 队列放置在模型调用前面。
- D
Increase maxTokens so fewer requests are needed.
增加 maxTokens 以减少需要的请求。
- E
Lower the temperature for more consistent outputs.
降低温度以获得更一致的输出。
解析 / Explanation
正确答案 / Answer: A,B
A. Backoff with jitter recovers from transient throttling/5xx without synchronized retry storms, directly improving reliability.
B. Spreading or failing over across Regions keeps the feature available during a regional impairment.
C. A queue in one Region is itself a regional single point of failure and adds latency to a synchronous, user-facing path, so it does not address regional resilience here.
D. Output length has no effect on resilience to errors or regional failure.
E. Temperature controls randomness, not reliability.
A. 带抖动的退避可以从瞬态节流/5xx 中恢复，无需同步重试风暴，直接提高可靠性。
B. 跨区域传播或故障转移可在区域故障期间保持该功能可用。
C. 一个区域中的队列本身就是一个区域单点故障，并且会增加面向用户的同步路径的延迟，因此它在这里不解决区域弹性问题。
D. 输出长度对错误或区域故障的恢复能力没有影响。
E. 温度控制随机性，而不是可靠性。
An interactive assistant must reduce the latency users perceive on long answers, WITHOUT changing which model is used and without a training step. Which choice is BEST?
交互式助手必须减少用户对长答案的感知延迟，而无需更改使用的模型，也无需进行训练步骤。哪个选择是最好的？
- A
Enable response streaming and turn on latency-optimized inference for the model.
启用响应流并打开模型的延迟优化推理。
- B
Distill to a smaller student model.
蒸馏为较小的学生模型。
- C
Buy Provisioned Throughput at high model units.
以高型号单位购买预置吞吐量。
- D
Switch to a model with a larger context window.
切换到具有更大上下文窗口的模型。
解析 / Explanation
正确答案 / Answer: A
A. Streaming shows tokens as they are generated (cutting perceived latency) and latency-optimized inference speeds generation, both without changing the model or training anything.
B. Distillation reduces latency but changes the model and requires a training step, which the scenario forbids.
C. Provisioned Throughput reduces throttling/queueing but does not lower the per-token generation time that dominates perceived latency on long answers.
D. A bigger context window does not reduce generation latency and can increase it.
A. 流式传输会显示生成的令牌（减少感知延迟），并且延迟优化的推理速度会加快生成速度，而无需更改模型或进行任何训练。
B. 蒸馏减少了延迟，但会改变模型并需要训练步骤，而该场景禁止这样做。
C. 预置吞吐量减少了限制/排队，但不会降低每个令牌的生成时间，而每个令牌的生成时间在长答案的感知延迟中占主导地位。
D. 更大的上下文窗口不会减少生成延迟，反而会增加延迟。
A team must roll out a new prompt version to production with the ability to revert instantly if quality regresses, measure both versions, and avoid redeploying application code. Which approach best fits Operational Excellence?
团队必须将新的即时版本投入生产，并且能够在质量下降时立即恢复、测量两个版本并避免重新部署应用程序代码。哪种方法最适合卓越运营？
- A
Manage the prompt in Amazon Bedrock Prompt Management with versions/variants, run automated evaluations, and switch or revert by changing the active version.
使用版本/变体管理 Amazon Bedrock Prompt Management 中的提示、运行自动评估以及通过更改活动版本进行切换或恢复。
- B
Hard-code the new prompt behind an application feature flag.
将新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, and split traffic.
部署两个 Lambda 函数，每个函数硬编码一个提示，并拆分流量。
- D
Raise the temperature so the new prompt is more robust.
提高温度，使新的提示更加有力。
解析 / Explanation
正确答案 / Answer: A
A. Managed prompt versions/variants plus automated evaluation let you compare, switch, and roll back without any code deploy.
B. A feature flag enables reverting but still requires a code deploy to ship the prompt and provides no managed versioning or evaluation.
C. This is custom, requires code deploys, and lacks managed prompt versioning and built-in evaluation.
D. Temperature does not provide rollout control, measurement, or rollback.
A. 托管提示版本/变体加上自动评估让您无需部署任何代码即可进行比较、切换和回滚。
B. 功能标志允许恢复，但仍然需要代码部署来发送提示，并且不提供托管版本控制或评估。
C. 这是自定义的，需要部署代码，并且缺乏托管提示版本控制和内置评估。
D. 温度不提供推出控制、测量或回滚。
A workload classifies tens of millions of documents per day in rolling waves; results must be available within about two hours of each submission (not real time). Cost must be minimized. Which option is MOST cost-effective?
工作负载每天对数以千万计的文档进行滚动分类；结果必须在每次提交后大约两小时内提供（非实时）。成本必须最小化。哪种选择最具成本效益？
- A
Amazon Bedrock batch inference jobs.
Amazon Bedrock 批量推理作业。
- B
On-demand real-time invocation for each document.
每个文档按需实时调用。
- C
Provisioned Throughput sized for the wave peaks.
针对波峰调整的预配置吞吐量。
- D
Amazon SageMaker Asynchronous Inference.
Amazon SageMaker 异步推理。
解析 / Explanation
正确答案 / Answer: A
A. The work is high-volume and latency-tolerant (about two hours), which is exactly what Bedrock batch inference is priced and designed for.
B. Per-document real-time calls work but are the most expensive option at this volume and are unnecessary given the two-hour tolerance.
C. Reserved capacity guarantees throughput but costs more than batch for latency-tolerant bulk work.
D. Asynchronous inference serves SageMaker-hosted models, not Bedrock foundation models, so it does not apply.
A. 这项工作量大且耐延迟（大约两个小时），这正是 Bedrock 批量推理的定价和设计目的。
B. 每个文档的实时调用是可行的，但在这个数量上是最昂贵的选择，并且考虑到两小时的容忍度，没有必要。
C. 预留容量保证了吞吐量，但对于耐延迟的批量工作来说，其成本高于批处理。
D. 异步推理服务于 SageMaker 托管的模型，而不是 Bedrock 基础模型，因此它不适用。
A team building a Bedrock RAG application wants AWS prescriptive best practices created specifically for generative AI concerns such as RAG accuracy, responsible AI, and model selection. Which resource is the BEST fit?
构建 Bedrock RAG 应用程序的团队希望专门针对生成式 AI 问题（例如 RAG 准确性、负责任的 AI 和模型选择）创建 AWS 规范性最佳实践。哪种资源最适合？
- A
The AWS Well-Architected Generative AI Lens.
AWS 架构完善的生成式 AI 镜头。
- B
The AWS Well-Architected Machine Learning Lens.
AWS 架构完善的机器学习镜头。
- C
The six core Well-Architected pillars alone.
仅六个核心架构良好的支柱。
- D
The AWS Well-Architected SaaS Lens.
AWS 架构完善的 SaaS 镜头。
解析 / Explanation
正确答案 / Answer: A
A. The Generative AI Lens provides design principles and best practices tailored to generative AI workloads, including RAG and responsible AI.
B. The ML Lens covers the traditional ML lifecycle and is not focused on generative-AI-specific concerns like RAG accuracy and responsible AI for FMs.
C. The core pillars are general guidance; the GenAI-specific practices live in the Generative AI Lens.
D. The SaaS Lens targets SaaS business/architecture concerns, not generative AI design.
A. 生成式 AI 镜头提供了针对生成式 AI 工作负载（包括 RAG 和负责任的 AI）量身定制的设计原则和最佳实践。
B. ML Lens 涵盖了传统的 ML 生命周期，并不专注于生成 AI 特定的问题，例如 RAG 准确性和 FM 的负责任的 AI。
C. 核心支柱是一般性指导； GenAI 的特定实践存在于生成式 AI 镜头中。
D. SaaS Lens 针对的是 SaaS 业务/架构问题，而不是生成式 AI 设计。
A RAG chatbot includes a 6,000-token static policy preamble in every request. The team wants to cut the cost and latency of that repeated context with minimal change and using the SAME model. Which option is BEST?
RAG 聊天机器人在每个请求中都包含 6,000 个令牌的静态策略前导码。该团队希望通过最小的更改并使用相同的模型来降低重复上下文的成本和延迟。哪个选项最好？
- A
Enable prompt caching for the static preamble so the repeated prefix is reused across requests.
为静态前导码启用提示缓存，以便在请求之间重复使用重复的前缀。
- B
Fine-tune the model so the policy is implicit and can be dropped from prompts.
微调模型，使策略是隐式的，并且可以从提示中删除。
- C
Store the preamble in a knowledge base and retrieve it each call.
将前导码存储在知识库中并在每次调用时检索它。
- D
Reduce maxTokens on the response.
减少响应的 maxTokens。
解析 / Explanation
正确答案 / Answer: A
A. Prompt caching reuses the unchanging prefix, reducing both billed tokens and latency for the repeated context with no model change.
B. Fine-tuning changes the model, adds maintenance, and is not a minimal change; it also makes policy updates harder.
C. Retrieval still injects the preamble tokens every call and adds a retrieval step, so it does not cut the repeated-token cost.
D. Lowering output length does not reduce the cost of the repeated input preamble.
A. 提示缓存重用不变的前缀，在不更改模型的情况下减少重复上下文的计费令牌和延迟。
B. 微调改变了模型，增加了维护，并不是一个最小的改变；这也使得政策更新变得更加困难。
C. 检索仍然会在每次调用时注入前导码令牌并添加检索步骤，因此它不会降低重复令牌成本。
D. 降低输出长度并不会降低重复输入前导码的成本。
A customer-facing RAG assistant occasionally produces confident but unsupported answers. The team wants to reduce hallucinations while keeping a good user experience, aligned with the GenAI Lens responsible-AI guidance. Which combination is BEST?
面向客户的 RAG 助理有时会给出自信但不受支持的答案。该团队希望减少幻觉，同时保持良好的用户体验，并与 GenAI Lens 负责的人工智能指南保持一致。哪种组合最好？
- A
Add a Guardrails contextual grounding check and instruct the model to answer only from the retrieved context and to say it does not know when the answer is absent.
添加 Guardrails 上下文接地检查，并指示模型仅从检索到的上下文中进行回答，并表示它不知道答案何时不存在。
- B
Set temperature to 0.
将温度设置为 0。
- C
Switch to a larger model.
切换到更大的型号。
- D
Add a denied-topics policy for the subject area.
为主题区域添加拒绝主题策略。
解析 / Explanation
正确答案 / Answer: A
A. Combining a grounding check with an explicit grounding/abstention instruction is defense-in-depth that targets ungrounded answers directly.
B. Lower randomness alone does not make an answer grounded in the sources; the model can still confidently fabricate.
C. A larger model still hallucinates without grounding controls; size is not a grounding mechanism.
D. Denied topics block whole subjects; they do not detect or prevent ungrounded answers within an allowed subject.
A. 将接地检查与明确的接地/弃权指令相结合是直接针对不接地气答案的深度防御。
B. 仅靠较低的随机性并不能得出基于来源的答案；该模型仍然可以放心地制造。
C. 较大的模型在没有接地控制的情况下仍然会产生幻觉；尺寸并不是一个接地机制。
D. 被拒绝的主题会阻塞整个主题；它们不会检测或阻止允许主题内的无根据答案。
A bank's assistant must always cite the LATEST internal regulations (which change weekly) AND must CONSISTENTLY emit a strict report template. Which approach meets BOTH with the least ongoing effort?
银行助理必须始终引用最新的内部规定（每周更改），并且必须始终如一地发布严格的报告模板。哪种方法可以用最少的持续努力满足两者？
- A
Use RAG for the frequently changing regulations and fine-tune the model for the fixed report template/style.
对于经常变化的法规使用 RAG，并针对固定的报告模板/样式微调模型。
- B
Fine-tune the model every week on both the regulations and the template.
每周根据规则和模板对模型进行微调。
- C
Use continued pre-training on the regulations and a prompt for the template.
继续进行有关规定的预培训和模板提示。
- D
Use RAG only, relying on a detailed prompt to enforce the template.
仅使用 RAG，依靠详细提示来强制执行模板。
解析 / Explanation
正确答案 / Answer: A
A. RAG keeps regulatory content current without retraining, while fine-tuning reliably enforces a rigid output format; each technique handles the part it is best at.
B. Weekly fine-tuning is operationally heavy and still lags between runs; RAG handles freshness far more cheaply.
C. Continued pre-training is for absorbing domain language from large corpora, is heavy, and does not keep weekly content current.
D. RAG fixes freshness, but prompt-only formatting is not reliably consistent, which the scenario explicitly requires.
A. RAG 无需重新培训即可保持监管内容最新，同时微调可靠地强制执行严格的输出格式；每种技术都处理它最擅长的部分。
B. 每周微调操作繁重，并且在运行之间仍然存在滞后； RAG handles freshness far more cheaply.
C. 持续的预训练是为了从大型语料库中吸收领域语言，训练量很大，并且不能保持每周的内容最新。
D. RAG 修复了新鲜度，但仅提示格式并不可靠一致，而场景明确要求这一点。
A team has a very large corpus of UNLABELED specialized legal text and wants the base model to internalize the domain's vocabulary and patterns. Which customization method is appropriate?
一个团队拥有一个非常大的未标记的专业法律文本语料库，并且希望基本模型能够内化该领域的词汇和模式。哪种定制方法合适？
- A
Continued pre-training on the unlabeled corpus.
继续在未标记的语料库上进行预训练。
- B
Supervised fine-tuning.
监督微调。
- C
Retrieval-Augmented Generation.
检索增强生成。
- D
Model distillation.
模型蒸馏。
解析 / Explanation
正确答案 / Answer: A
A. Continued pre-training adapts a model to domain language using large volumes of unlabeled text.
B. Fine-tuning requires labeled prompt/response pairs; the corpus is unlabeled, so it does not apply.
C. RAG retrieves text at inference time but does not make the model internalize domain language.
D. Distillation compresses a larger model into a smaller one; it is not a method for absorbing a domain corpus.
A. 持续的预训练使用大量未标记的文本使模型适应领域语言。
B. 微调需要标记提示/响应对；该语料库未标记，因此不适用。
C. RAG 在推理时检索文本，但不会使模型内化领域语言。
D. 蒸馏将较大的模型压缩为较小的模型；它不是吸收领域语料库的方法。
Latency and cost of a large model are too high, but the team must KEEP its accuracy on their specific task and can accept an upfront training step. Which option BEST fits?
大型模型的延迟和成本太高，但团队必须保持其特定任务的准确性，并且可以接受前期训练步骤。哪个选项最适合？
- A
Model distillation, training a smaller student from the large teacher on the task.
模型蒸馏，从大老师那里训练较小的学生来完成任务。
- B
Enable latency-optimized inference on the large model.
在大型模型上启用延迟优化推理。
- C
Purchase Provisioned Throughput for the large model.
购买大型模型的预配置吞吐量。
- D
Lower top-k to make the large model cheaper.
降低 top-k 可以使大型模型更便宜。
解析 / Explanation
正确答案 / Answer: A
A. Distillation produces a smaller, cheaper, faster model that retains much of the teacher's task accuracy, and it explicitly involves an upfront training step.
B. Latency-optimized inference speeds responses but does not meaningfully cut the large model's cost and involves no training step.
C. Provisioned Throughput stabilizes capacity but does not reduce the per-call cost or size of the large model.
D. Sampling parameters do not change model cost or latency materially.
A. 蒸馏产生了一个更小、更便宜、更快的模型，保留了教师任务的大部分准确性，并且它明确涉及预先训练步骤。
B. 延迟优化推理可加快响应速度，但不会显着降低大型模型的成本，并且不涉及任何训练步骤。
C. 预置吞吐量可以稳定容量，但不会降低每次调用的成本或大型模型的大小。
D. 采样参数不会显着改变模型成本或延迟。
Throughput must increase for a steady, predictable, high baseline load running 24/7, at the lowest per-token cost, using the same model and API. Which option is BEST?
必须使用相同的模型和 API，以最低的每个代币成本，增加吞吐量，以实现 24/7 运行的稳定、可预测、高基线负载。哪个选项最好？
- A
Provisioned Throughput with a committed term.
在承诺期限内配置吞吐量。
- B
Cross-Region inference.
跨区域推理。
- C
Bedrock batch inference.
基岩批量推理。
- D
On-demand inference with client retries.
通过客户端重试进行按需推理。
解析 / Explanation
正确答案 / Answer: A
A. For steady 24/7 high volume, a committed Provisioned Throughput model unit gives the lowest per-token cost and guaranteed capacity.
B. Cross-Region inference adds throughput but is on-demand priced, so it is not the lowest per-token cost for steady 24/7 load.
C. Batch inference is offline and cannot serve steady real-time throughput.
D. On-demand at constant high volume costs more and risks throttling versus committed reserved capacity.
A. 为了实现稳定的 24/7 高容量，承诺的预配置吞吐量模型单元可提供最低的每个代币成本和有保证的容量。
B. 跨区域推理增加了吞吐量，但按需求定价，因此它不是稳定 24/7 负载的最低每个令牌成本。
C. 批量推理是离线的，无法提供稳定的实时吞吐量。
D. 与承诺的预留容量相比，持续高容量的按需成本更高，并且存在节流风险。
To make a model's outputs maximally deterministic for snapshot tests, which TWO inference settings TOGETHER are most effective? (Select TWO.)
为了使模型的输出对于快照测试具有最大确定性，哪两个推理设置组合起来最有效？ （选择两项。）
- A
Set temperature to 0.
将温度设置为 0。
- B
Constrain sampling with a low top-p (and avoid a large top-k).
使用较低的 top-p 约束采样（并避免较大的 top-k）。
- C
Use a high temperature with a low maxTokens.
使用较高的温度和较低的 maxTokens。
- D
Add several stop sequences.
添加几个停止序列。
- E
Increase top-k substantially.
大幅增加top-k。
解析 / Explanation
正确答案 / Answer: A,B
A. A temperature of 0 makes sampling effectively greedy, the primary lever for deterministic output.
B. Narrowing the nucleus to the most probable tokens removes residual variability that temperature alone may leave.
C. High temperature maximizes randomness; capping length does not make content deterministic.
D. Stop sequences control where generation ends, not how deterministic the tokens are.
E. A larger top-k widens the candidate pool and increases variability.
A. 温度为 0 使得采样实际上变得贪婪，这是确定性输出的主要杠杆。
B. 将核缩小到最可能的标记可以消除仅温度可能留下的残余变异性。
C. 高温使随机性最大化；限制长度不会使内容具有确定性。
D. 停止序列控制生成的结束位置，而不是令牌的确定性。
E. 较大的 top-k 会扩大候选池并增加可变性。
A pipeline expects the model to stop generating exactly when it produces the marker END_JSON, and that marker must not appear in the output to keep downstream parsing clean. Which solution is BEST?
管道期望模型在生成标记 END_JSON 时停止生成，并且该标记不得出现在输出中以保持下游解析干净。哪种解决方案最好？
- A
Configure END_JSON as a stop sequence in the inference request.
将 END_JSON 配置为推理请求中的停止序列。
- B
Set maxTokens to an estimated length.
将 maxTokens 设置为估计长度。
- C
Lower the temperature.
降温。
- D
Instruct the model in the prompt to stop after END_JSON.
指示提示中的模型在 END_JSON 之后停止。
解析 / Explanation
正确答案 / Answer: A
A. A stop sequence halts generation when the phrase is produced and the stop text itself is not included in the returned output, which keeps parsing clean.
B. A length guess does not stop at the marker and risks cutting content or leaving the marker in.
C. Temperature does not control stopping at a specific phrase.
D. Prompt instructions are not a reliable stopping mechanism and would leave the marker in the text.
A. 当生成短语并且停止文本本身不包含在返回的输出中时，停止序列会停止生成，从而保持解析干净。
B. 长度猜测不会在标记处停止，并且存在剪切内容或留下标记的风险。
C. 温度不控制在特定阶段停止。
D. 提示指令不是可靠的停止机制，并且会在文本中留下标记。
A high-volume pipeline must caption product IMAGES with short text at the lowest cost and latency while still supporting vision. Which Amazon Nova model is BEST?
高容量管道必须以最低的成本和延迟为产品图像添加短文本字幕，同时仍然支持视觉。哪种 Amazon Nova 型号最好？
- A
Amazon Nova Lite.
亚马逊 Nova Lite。
- B
Amazon Nova Micro.
亚马逊新星微。
- C
Amazon Nova Pro.
亚马逊 Nova Pro。
- D
Amazon Nova Canvas.
亚马逊新星画布。
解析 / Explanation
正确答案 / Answer: A
A. Nova Lite is the low-cost multimodal model, the cheapest/fastest Nova option that still understands images.
B. Nova Micro is the lowest-cost model overall but is text-only and cannot process images.
C. Nova Pro understands images but at higher cost and latency than this simple, high-volume captioning needs.
D. Nova Canvas generates images; it does not caption or understand existing images.
A. Nova Lite 是低成本的多模式模型，是仍能理解图像的最便宜/最快的 Nova 选项。
B. Nova Micro 是总体成本最低的型号，但仅支持文本且无法处理图像。
C. Nova Pro 可以理解图像，但与这种简单的大容量字幕需求相比，成本和延迟更高。
D. Nova Canvas 生成图像；它不提供字幕或理解现有图像。
A team must cut vector-storage cost for a TEXT-ONLY semantic search while keeping good recall, with minimal re-architecture. Which choice is BEST?
团队必须削减纯文本语义搜索的向量存储成本，同时保持良好的召回率，并尽可能减少重新架构。哪个选择是最好的？
- A
Use Amazon Titan Text Embeddings V2 and select a smaller output dimension.
使用 Amazon Titan Text Embeddings V2 并选择较小的输出维度。
- B
Switch to Amazon Titan Multimodal Embeddings.
切换到 Amazon Titan 多模式嵌入。
- C
Increase the embedding dimension to improve accuracy.
增加嵌入尺寸以提高精度。
- D
Use a text-generation model to produce the embeddings.
使用文本生成模型来生成嵌入。
解析 / Explanation
正确答案 / Answer: A
A. Titan Text Embeddings V2 supports configurable, smaller dimensions that cut index size and cost while keeping acceptable recall, with no architecture change.
B. Multimodal embeddings target image+text retrieval and do not reduce storage for a text-only search.
C. Larger dimensions increase storage and cost, the opposite of the goal.
D. Generative models are not embedding models and would not provide a proper vector index.
A. Titan Text Embeddings V2 支持可配置的较小尺寸，可以减少索引大小和成本，同时保持可接受的召回率，而无需更改架构。
B. 多模态嵌入的目标是图像+文本检索，并且不会减少纯文本搜索的存储空间。
C. 较大的尺寸会增加存储和成本，这与目标相反。
D. 生成模型不是嵌入模型，并且不会提供适当的向量索引。
A team must compare two models on SUMMARY quality (coherence and relevance) over thousands of documents, automatically, with NO human raters and NO reference summaries available. Which evaluation approach fits?
团队必须自动比较数千份文档中两个模型的摘要质量（连贯性和相关性），无需人工评分者和可用的参考摘要。哪种评估方法适合？
- A
Amazon Bedrock model evaluation with an LLM-as-a-judge evaluator.
由 LLM 法官评估员进行 Amazon Bedrock 模型评估。
- B
Bedrock evaluation using reference-based metrics such as BERTScore or exact match.
使用基于参考的指标（例如 BERTScore 或精确匹配）进行基岩评估。
- C
A human evaluation workflow.
人工评估工作流程。
- D
Compare the two models' CloudWatch latency metrics.
比较两个模型的 CloudWatch 延迟指标。
解析 / Explanation
正确答案 / Answer: A
A. LLM-as-a-judge scores open-ended quality dimensions at scale without humans and without needing reference answers.
B. Reference-based metrics require ground-truth reference summaries, which are not available here.
C. Human evaluation is excluded by the no-human-raters constraint.
D. Latency does not measure summary coherence or relevance.
A. 法学硕士法官可以在没有人工参与且不需要参考答案的情况下对开放式质量维度进行大规模评分。
B. 基于参考的指标需要真实的参考摘要，但此处不提供。
C. 人类评估被无人类评估者约束排除在外。
D. 延迟并不衡量摘要的一致性或相关性。
A team must build ONE codebase that does multi-turn chat WITH tool calling across both Anthropic Claude and Amazon Nova models, minimizing model-specific request/response code. Which approach is BEST?
A team must build ONE codebase that does multi-turn chat WITH tool calling across both Anthropic Claude and Amazon Nova models, minimizing model-specific request/response code.哪种方法最好？
- A
Use the Amazon Bedrock Converse API.
使用 Amazon Bedrock Converse API。
- B
Call InvokeModel with each model's native JSON body.
使用每个模型的本机 JSON 正文调用 InvokeModel。
- C
Use InvokeModelWithResponseStream per model.
每个模型使用 InvokeModelWithResponseStream。
- D
Build a custom abstraction layer over each model's API.
在每个模型的 API 上构建自定义抽象层。
解析 / Explanation
正确答案 / Answer: A
A. Converse provides a unified, model-agnostic interface for multi-turn messages and tool use across supported models, minimizing per-model code.
B. InvokeModel requires model-specific payloads and parsing, which is exactly the per-model code to avoid.
C. This still needs model-specific bodies and only adds streaming, not unification.
D. A custom layer reinvents what Converse already standardizes.
A. Converse 为跨受支持模型的多轮消息和工具使用提供了一个与模型无关的统一界面，从而最大限度地减少了每个模型的代码。
B. InvokeModel 需要特定于模型的有效负载和解析，这正是要避免的每个模型代码。
C. 这仍然需要特定于模型的主体，并且仅添加流，而不是统一。
D. 定制层重塑了 Converse 已经标准化的内容。
A company fine-tuned open-weights Llama models OUTSIDE AWS (a supported architecture) and wants them callable through the Amazon Bedrock runtime API, on demand, with no servers to run. Which option is correct?
一家公司对 AWS（受支持的架构）之外的开放权重 Llama 模型进行了微调，并希望它们可以根据需要通过 Amazon Bedrock 运行时 API 进行调用，而无需运行服务器。哪个选项是正确的？
- A
Amazon Bedrock Custom Model Import.
Amazon Bedrock 自定义模型导入。
- B
Deploy the weights to a SageMaker real-time endpoint.
将权重部署到 SageMaker 实时端点。
- C
Buy Provisioned Throughput on a Bedrock base model.
购买基岩基础模型上的预配置吞吐量。
- D
Run continued pre-training in Bedrock.
在基岩中继续进行预训练。
解析 / Explanation
正确答案 / Answer: A
A. Custom Model Import brings supported external weights into Bedrock so they are invoked through the standard on-demand Bedrock APIs with no managed infrastructure.
B. A SageMaker endpoint can serve the model but is not the Bedrock runtime API and requires managing the endpoint.
C. Provisioned Throughput reserves capacity for Bedrock models; it does not serve your own imported weights.
D. Continued pre-training customizes a Bedrock base model; it does not import your externally trained weights.
A. 自定义模型导入将受支持的外部权重引入 Bedrock，以便通过标准的按需 Bedrock API 调用它们，无需托管基础设施。
B. SageMaker 端点可以为模型提供服务，但不是 Bedrock 运行时 API，需要管理端点。
C. 预配置吞吐量为基岩模型保留容量；它不服务于您自己导入的权重。
D. 持续的预训练定制了Bedrock基础模型；它不会导入您外部训练的权重。
A team wants to grow throughput using cross-Region inference, but compliance requires that all inference data remain within one approved Region. What is the CORRECT stance?
团队希望使用跨区域推理来提高吞吐量，但合规性要求所有推理数据都保留在一个批准的区域内。正确的立场是什么？
- A
Do not use a cross-Region profile that can route outside the approved Region; stay on in-Region on-demand or Provisioned Throughput.
请勿使用可以路由到批准区域之外的跨区域配置文件；保持区域内按需或预置吞吐量。
- B
Enable cross-Region inference; the data never leaves the source Region.
启用跨Region推理；数据永远不会离开源区域。
- C
Enable cross-Region inference; it always keeps data within the source Region's country.
启用跨Region推理；它始终将数据保存在源区域的国家/地区内。
- D
Switch to batch inference to keep data in Region.
切换到批量推理以将数据保留在 Region 中。
解析 / Explanation
正确答案 / Answer: A
A. Cross-Region inference can route requests to other Regions in the geography, which would violate single-Region data residency, so it must be avoided for this requirement.
B. Cross-Region inference can process requests in other Regions, so this statement is incorrect.
C. Routing targets depend on the profile's geography and are not guaranteed to stay in one country/Region.
D. Batch versus real-time is unrelated to the residency guarantee and does not address the throughput requirement correctly.
A. 跨区域推理可以将请求路由到地理上的其他区域，这会违反单区域数据驻留，因此必须避免此要求。
B. 跨Region推理可以处理其他Region的请求，所以这种说法是不正确的。
C. 路由目标取决于配置文件的地理位置，并且不保证位于一个国家/地区。
D. 批量与实时与驻留保证无关，并且不能正确满足吞吐量要求。
Users upload SCREENSHOTS OF CHARTS and ask analytical questions about the trends shown. Which approach BEST answers them?
用户上传图表屏幕截图并就所显示的趋势提出分析问题。哪种方法最能回答他们的问题？
- A
Send the image to a Bedrock multimodal foundation model via the Converse API.
通过 Converse API 将图像发送到 Bedrock 多模式基础模型。
- B
Use Amazon Textract to extract the chart.
使用 Amazon Textract 提取图表。
- C
Use Amazon Rekognition to analyze the image.
使用 Amazon Rekognition 分析图像。
- D
Use Amazon Titan Multimodal Embeddings.
使用 Amazon Titan 多模式嵌入。
解析 / Explanation
正确答案 / Answer: A
A. A multimodal FM can interpret the chart's visuals and answer natural-language analytical questions about the trends.
B. Textract extracts text and tables from documents but does not reason about a chart's visual trends.
C. Rekognition returns labels/objects, not analytical answers about chart trends.
D. Multimodal embeddings support retrieval/similarity, not question answering about an image.
A. 多模式 FM 可以解释图表的视觉效果并回答有关趋势的自然语言分析问题。
B. Textract 从文档中提取文本和表格，但不会推断图表的视觉趋势。
C. Rekognition 返回标签/对象，而不是有关图表趋势的分析答案。
D. 多模态嵌入支持检索/相似性，而不是关于图像的问答。
A Bedrock foundation model must process a very large OFFLINE job over millions of prompts at the lowest cost; turnaround of tens of minutes is fine. Which option is BEST?
Bedrock 基础模型必须以最低的成本处理数百万条提示的超大型离线作业；几十分钟的周转就可以了。哪个选项最好？
- A
Amazon Bedrock batch inference (a model invocation job).
Amazon Bedrock 批量推理（模型调用作业）。
- B
Amazon SageMaker Asynchronous Inference.
Amazon SageMaker 异步推理。
- C
A loop of on-demand InvokeModel calls.
按需 InvokeModel 调用的循环。
- D
Provisioned Throughput.
预置吞吐量。
解析 / Explanation
正确答案 / Answer: A
A. Batch inference is purpose-built for large, latency-tolerant offline workloads on Bedrock FMs at the lowest cost.
B. Asynchronous inference serves SageMaker-hosted models, not Bedrock foundation models, so it does not apply.
C. Real-time calls in a loop are more expensive and slower for millions of prompts than a batch job.
D. Reserved online capacity is not the cheapest way to run a one-off offline bulk job.
A. 批量推理专为以最低成本在 Bedrock FM 上处理大型、延迟容忍的离线工作负载而构建。
B. 异步推理服务于 SageMaker 托管的模型，而不是 Bedrock 基础模型，因此它不适用。
C. 对于数百万个提示，循环中的实时调用比批处理作业更昂贵且更慢。
D. 预留的在线容量并不是运行一次性离线批量作业的最便宜的方式。
A team has only about 30 examples of a desired behavior and needs better adherence quickly; collecting thousands of labeled examples is not feasible. Which approach is BEST?
一个团队只有大约 30 个所需行为的例子，需要快速更好地遵守；收集数千个带标签的示例是不可行的。哪种方法最好？
- A
Use few-shot prompting that includes the example demonstrations.
使用包含示例演示的少量提示。
- B
Fine-tune the model on the 30 examples.
在 30 个示例上微调模型。
- C
Run continued pre-training.
继续进行预训练。
- D
Distill the model.
提取模型。
解析 / Explanation
正确答案 / Answer: A
A. Few-shot prompting leverages a handful of examples immediately and is the right tool when labeled data is scarce.
B. Roughly 30 examples are generally too few to fine-tune effectively and risk overfitting with little gain.
C. Continued pre-training needs a large unlabeled corpus, not 30 examples.
D. Distillation addresses cost/latency, not teaching a behavior from a few examples.
A. 少样本提示可以立即利用少量示例，并且在标记数据稀缺时是正确的工具。
B. 大约 30 个示例通常太少，无法进行有效微调，并且存在过度拟合而收益甚微的风险。
C. 持续的预训练需要大量未标记的语料库，而不是 30 个示例。
D. 蒸馏解决的是成本/延迟，而不是通过几个例子来教授行为。
A summarizer sometimes truncates mid-sentence, but the team also wants to prevent runaway, overly long outputs. Which TWO settings TOGETHER best address both? (Select TWO.)
摘要生成器有时会截断句子中间的内容，但团队也希望防止出现失控、过长的输出。哪两个设置组合在一起最能同时解决这两个问题？ （选择两项。）
- A
Increase maxTokens enough to fit a complete answer.
增加 maxTokens 足以适应完整的答案。
- B
Add a stop sequence at the natural end-of-answer marker.
在自然答案结束标记处添加停止序列。
- C
Lower the temperature.
降温。
- D
Increase top-k.
增加top-k。
- E
Enlarge the model's context window.
放大模型的上下文窗口。
解析 / Explanation
正确答案 / Answer: A,B
A. Mid-sentence truncation means the output is hitting the token cap, so raising maxTokens to a sufficient size lets answers complete.
B. A stop sequence ends generation cleanly at the intended boundary, preventing runaway length.
C. Temperature affects randomness, not truncation or maximum length.
D. Top-k changes token diversity, not output length.
E. The context window governs input size, not how long the output may be.
A. 句子中间截断意味着输出达到了令牌上限，因此将 maxTokens 提高到足够的大小可以让答案完整。
B. 停止序列在预期边界处干净地结束生成，防止长度失控。
C. 温度影响随机性，而不影响截断或最大长度。
D. Top-k 改变 token 多样性，而不是输出长度。
E. 上下文窗口控制输入大小，而不是输出的长度。
Before fine-tuning on customer chat transcripts, a team must programmatically REMOVE PII from the free text across a large dataset. Which service is the right tool?
在对客户聊天记录进行微调之前，团队必须以编程方式从大型数据集中的自由文本中删除 PII。哪种服务是正确的工具？
- A
Amazon Comprehend PII detection and redaction.
Amazon Comprehend PII 检测和编辑。
- B
Amazon Macie.
亚马逊麦西.
- C
Amazon Bedrock Guardrails sensitive-information filter.
Amazon Bedrock Guardrails 敏感信息过滤器。
- D
Amazon Textract.
亚马逊文本。
解析 / Explanation
正确答案 / Answer: A
A. Comprehend detects and redacts PII entities within text, producing sanitized transcripts for training.
B. Macie discovers and classifies sensitive data across S3 but does not redact PII inside text for downstream training.
C. Guardrails apply at inference time to a model's prompts/responses; they are not a batch redaction tool for preparing a training corpus.
D. Textract extracts text from documents; it does not detect or redact PII.
A. Comprehend 检测并编辑文本中的 PII 实体，生成经过净化的训练记录。
B. Macie 发现 S3 中的敏感数据并对其进行分类，但不会编辑文本内的 PII 以进行下游训练。
C. 护栏在推理时应用于模型的提示/响应；它们不是用于准备训练语料库的批量编辑工具。
D. Textract 从文档中提取文本；它不会检测或编辑 PII。
A governance team must INVENTORY where PII resides across 500 Amazon S3 buckets before deciding how to process it. Which service fits BEST?
治理团队必须先清查 500 个 Amazon S3 存储桶中 PII 所在的位置，然后再决定如何处理它。哪种服务最适合？
- A
Amazon Macie.
亚马逊麦西.
- B
Amazon Comprehend.
亚马逊理解。
- C
Amazon Bedrock Guardrails.
亚马逊基岩护栏。
- D
Amazon S3 Inventory.
亚马逊 S3 库存。
解析 / Explanation
正确答案 / Answer: A
A. Macie is built to discover and classify sensitive data and PII across S3 at scale, producing the inventory the team needs.
B. Comprehend analyzes and redacts PII within individual text inputs but is not a bucket-scale discovery/classification inventory tool.
C. Guardrails govern model I/O at inference; they do not scan storage for sensitive data.
D. S3 Inventory lists objects and metadata but does not classify which contain PII.
A. Macie 旨在大规模发现和分类 S3 中的敏感数据和 PII，从而生成团队所需的清单。
B. 理解分析和编辑单个文本输入中的 PII，但不是桶级发现/分类清单工具。
C. Guardrails 控制推理时的模型 I/O；他们不会扫描存储中的敏感数据。
D. S3 Inventory 列出了对象和元数据，但未对包含 PII 的对象和元数据进行分类。
A team needs human-labeled instruction/response pairs with a managed workforce option and built-in quality controls (annotation consolidation). Which service is BEST?
团队需要人工标记的指令/响应对，以及受管理的劳动力选项和内置的质量控制（注释整合）。哪种服务最好？
- A
Amazon SageMaker Ground Truth.
Amazon SageMaker 地面真相。
- B
Raw Amazon Mechanical Turk tasks.
原始 Amazon Mechanical Turk 任务。
- C
Amazon SageMaker Data Wrangler.
Amazon SageMaker 数据牧马人。
- D
Amazon Comprehend custom classification.
Amazon Comprehend 自定义分类。
解析 / Explanation
正确答案 / Answer: A
A. Ground Truth provides managed labeling workflows, workforce options, and quality features such as annotation consolidation.
B. Mechanical Turk supplies a crowd but lacks the managed labeling workflow, templates, and consolidation that Ground Truth adds.
C. Data Wrangler transforms tabular data; it does not run human annotation.
D. Comprehend automates classification; it is not a human-labeling service for creating new pairs.
A. Ground Truth 提供托管标签工作流程、劳动力选项和注释整合等质量功能。
B. Mechanical Turk 提供了大量服务，但缺乏 Ground Truth 所添加的托管标签工作流程、模板和整合。
C. Data Wrangler 转换表格数据；它不运行人工注释。
D. 理解自动分类；它不是用于创建新配对的人工标记服务。
Millions of raw JSON documents must be cleaned, normalized, and partitioned every night as a repeatable, serverless production pipeline with the least operational overhead. Which service is BEST?
每晚必须对数百万个原始 JSON 文档进行清理、规范化和分区，作为可重复、无服务器的生产管道，并且运营开销最少。哪种服务最好？
- A
AWS Glue serverless ETL jobs.
AWS Glue 无服务器 ETL 作业。
- B
Amazon SageMaker Data Wrangler.
Amazon SageMaker 数据牧马人。
- C
An Amazon EMR Spark cluster.
Amazon EMR Spark 集群。
- D
One AWS Lambda invocation per file.
每个文件一次 AWS Lambda 调用。
解析 / Explanation
正确答案 / Answer: A
A. Glue runs scheduled, serverless Spark ETL for large-scale cleaning/partitioning with no clusters to manage.
B. Data Wrangler is an interactive, notebook-oriented prep tool, not a hands-off serverless nightly production pipeline at this scale.
C. EMR is capable but requires provisioning and managing cluster capacity.
D. Per-file Lambda processing is hard to orchestrate at this scale and hits timeout limits.
A. Glue 运行计划的无服务器 Spark ETL，用于大规模清理/分区，无需管理集群。
B. Data Wrangler 是一种交互式、面向笔记本的准备工具，而不是这种规模的不干涉的无服务器夜间生产管道。
C. EMR 有能力，但需要配置和管理集群容量。
D. 按文件 Lambda 处理很难在这种规模下进行编排，并且会达到超时限制。
A developer must provide a SUPERVISED fine-tuning dataset to Amazon Bedrock. What is the correct format and location?
开发人员必须向 Amazon Bedrock 提供受监督的微调数据集。正确的格式和位置是什么？
- A
A JSONL file of prompt/completion (or messages) examples in Amazon S3 matching the model's schema.
Amazon S3 中与模型架构匹配的提示/完成（或消息）示例的 JSONL 文件。
- B
A CSV export of raw documents in S3.
S3 中原始文档的 CSV 导出。
- C
A folder of PDFs in S3.
S3 中的 PDF 文件夹。
- D
An Amazon DynamoDB table of examples.
Amazon DynamoDB 示例表。
解析 / Explanation
正确答案 / Answer: A
A. Bedrock customization reads labeled examples as JSONL records from S3 in the required schema.
B. Raw documents are not labeled prompt/response pairs; this is not a supervised fine-tuning dataset.
C. An unstructured PDF corpus suits RAG ingestion, not supervised fine-tuning.
D. Fine-tuning reads from S3 in the required file format, not from DynamoDB.
A. Bedrock 自定义将标记的示例从所需架构中的 S3 中读取为 JSONL 记录。
B. 原始文档没有标记提示/响应对；这不是一个有监督的微调数据集。
C. 非结构化 PDF 语料库适合 RAG 摄取，而不是监督微调。
D. 微调以所需的文件格式从 S3 读取，而不是从 DynamoDB 读取。
A team must measure representation imbalance in the TRAINING DATASET before any training begins. Which capability is correct?
在任何训练开始之前，团队必须测量训练数据集中的代表性不平衡。哪种能力是正确的？
- A
Amazon SageMaker Clarify pre-training bias metrics.
Amazon SageMaker 澄清预训练偏差指标。
- B
Amazon SageMaker Clarify post-training bias metrics.
Amazon SageMaker 澄清训练后偏差指标。
- C
Amazon SageMaker Model Monitor.
Amazon SageMaker 模型监视器。
- D
Amazon Comprehend.
亚马逊理解。
解析 / Explanation
正确答案 / Answer: A
A. Clarify computes pre-training bias metrics directly on the dataset before training.
B. Post-training metrics evaluate a trained model's predictions, not the dataset before training.
C. Model Monitor detects drift on a deployed endpoint, not dataset bias before training.
D. Comprehend performs NLP tasks, not dataset bias measurement.
A. Clarify 在训练前直接在数据集上计算训练前偏差指标。
B. 训练后指标评估训练后模型的预测，而不是训练前的数据集。
C. 模型监视器检测已部署端点上的漂移，而不是训练前的数据集偏差。
D. Comprehend 执行 NLP 任务，而不是数据集偏差测量。
A governance team must publish a standardized, human-readable record of each model's intended use, training data, evaluation, and risk rating for audits. Which feature is BEST?
治理团队必须发布每个模型的预期用途、培训数据、评估和审计风险评级的标准化、人类可读的记录。哪个功能最好？
- A
Amazon SageMaker Model Cards.
Amazon SageMaker 模型卡。
- B
Amazon SageMaker ML Lineage Tracking.
Amazon SageMaker ML 沿袭跟踪。
- C
AWS CloudTrail.
AWS CloudTrail。
- D
AWS Config.
AWS 配置。
解析 / Explanation
正确答案 / Answer: A
A. Model Cards capture intended use, training details, evaluation, and risk in a standardized, versioned, auditable document.
B. Lineage tracking records artifact/step graphs automatically; it is not the human-readable governance document of intended use and risk.
C. CloudTrail logs API activity, not model governance documentation.
D. Config tracks resource configuration compliance, not model documentation.
A. 模型卡以标准化、版本化、可审核的文档捕获预期用途、培训细节、评估和风险。
B. 沿袭跟踪自动记录工件/步骤图；它不是关于预期用途和风险的人类可读的治理文件。
C. CloudTrail 记录 API 活动，而不是模型治理文档。
D. Config 跟踪资源配置合规性，而不是模型文档。
A team must AUTOMATICALLY track which dataset version and which processing job produced a given model artifact, as a queryable lineage graph. Which capability is correct?
团队必须自动跟踪哪个数据集版本以及哪个处理作业生成了给定的模型工件，作为可查询的谱系图。哪种能力是正确的？
- A
Amazon SageMaker ML Lineage Tracking.
Amazon SageMaker ML 沿袭跟踪。
- B
Amazon SageMaker Model Cards.
Amazon SageMaker 模型卡。
- C
AWS Glue Data Catalog.
AWS Glue 数据目录。
- D
AWS CloudTrail.
AWS CloudTrail。
解析 / Explanation
正确答案 / Answer: A
A. ML Lineage Tracking automatically records the entities and relationships (datasets, jobs, models) as a queryable lineage graph.
B. Model Cards are a manually authored governance document, not an automatic artifact-lineage graph.
C. The Data Catalog inventories data sources; it does not track model-artifact lineage.
D. CloudTrail records API calls, not a model-artifact lineage graph.
A. ML Lineage Tracking 自动将实体和关系（数据集、作业、模型）记录为可查询的谱系图。
B. 模型卡是手动编写的治理文档，而不是自动的工件谱系图。
C. 数据目录清单数据源；它不跟踪模型-工件谱系。
D. CloudTrail 记录 API 调用，而不是模型工件沿袭图。
A team is preparing a fine-tuning dataset and wants to minimize overfitting and train/test leakage. Which TWO steps best support this? (Select TWO.)
一个团队正在准备微调数据集，并希望最大限度地减少过度拟合和训练/测试泄漏。哪两个步骤最能支持这一点？ （选择两项。）
- A
Deduplicate near-duplicate examples.
对接近重复的示例进行重复数据删除。
- B
Create a held-out test split that does not overlap the training data.
创建不与训练数据重叠的保留测试分割。
- C
Oversample the majority class to grow the dataset.
对多数类进行过采样以增长数据集。
- D
Include the evaluation questions in the training set for coverage.
将评估问题包含在训练集中以进行覆盖。
- E
Increase the number of training epochs.
增加训练纪元数。
解析 / Explanation
正确答案 / Answer: A,B
A. Removing duplicates prevents the model from overweighting repeated samples and reduces overfitting.
B. A non-overlapping evaluation split prevents leakage and gives an honest generalization estimate.
C. Oversampling the majority worsens imbalance and encourages overfitting.
D. Putting test items into training is textbook data leakage.
E. Epoch count is a training hyperparameter, not a data-preparation step, and more epochs can overfit.
A. 删除重复项可以防止模型过度加权重复样本并减少过度拟合。
B. 非重叠的评估分割可以防止泄漏并给出诚实的概括估计。
C. 对大多数人进行过度采样会加剧不平衡并鼓励过度拟合。
D. 将测试项目投入训练就是教科书数据泄露。
E. 纪元计数是一个训练超参数，而不是数据准备步骤，更多纪元可能会过度拟合。
A Bedrock knowledge base needs a fully managed vector store for HIGH query-per-second, LOW-latency similarity search with no capacity planning. Which option is BEST?
Bedrock 知识库需要完全托管的矢量存储，以实现每秒高查询、低延迟的相似性搜索，无需进行容量规划。哪个选项最好？
- A
Amazon OpenSearch Serverless.
亚马逊 OpenSearch 无服务器。
- B
Amazon Aurora PostgreSQL with pgvector.
带有 pgvector 的 Amazon Aurora PostgreSQL。
- C
Amazon S3 Vectors.
亚马逊 S3 向量。
- D
Amazon Neptune Analytics.
亚马逊海王星分析。
解析 / Explanation
正确答案 / Answer: A
A. OpenSearch Serverless is the natively integrated, auto-scaling managed vector store suited to frequent, low-latency queries without capacity planning.
B. Aurora pgvector is supported but you size and manage the database instance; it is not the auto-scaling serverless choice for high QPS.
C. S3 Vectors optimizes cost for infrequent queries at huge scale, not high-QPS low-latency search.
D. Neptune Analytics powers GraphRAG over relationships, not general high-QPS vector search.
A. OpenSearch Serverless 是原生集成、自动扩展的托管矢量存储，适合频繁、低延迟的查询，无需进行容量规划。
B. 支持 Aurora pgvector，但您可以调整和管理数据库实例；它不是高 QPS 的自动扩展无服务器选择。
C. S3 Vectors 优化了大规模不频繁查询的成本，而不是高 QPS 低延迟搜索的成本。
D. Neptune Analytics 通过关系而非一般的高 QPS 向量搜索为 GraphRAG 提供支持。
A corpus of billions of embeddings is queried only a few times per month for occasional investigations. The team wants the lowest-cost durable vector storage with no capacity to manage. Which option is BEST?
包含数十亿嵌入的语料库每月仅查询几次以进行偶尔的调查。该团队希望获得成本最低且无需管理容量的持久矢量存储。哪个选项最好？
- A
Amazon S3 Vectors.
亚马逊 S3 向量。
- B
Amazon OpenSearch Serverless.
亚马逊 OpenSearch 无服务器。
- C
Amazon Aurora PostgreSQL with pgvector.
带有 pgvector 的 Amazon Aurora PostgreSQL。
- D
Amazon DynamoDB.
亚马逊 DynamoDB。
解析 / Explanation
正确答案 / Answer: A
A. S3 Vectors gives serverless, pay-for-use vector storage and search at massive scale, ideal for infrequent queries.
B. OpenSearch Serverless keeps warm capacity you pay for, which is wasteful for queries a few times a month.
C. Aurora requires running instances and is not the lowest-cost option for rarely queried, very large vector sets.
D. DynamoDB has no native vector similarity search.
A. S3 Vectors 提供无服务器、按使用付费的大规模矢量存储和搜索，非常适合不频繁的查询。
B. OpenSearch Serverless 可保持您付费的热容量，这对于每月几次查询来说是浪费的。
C. Aurora 需要运行实例，对于很少查询的非常大的向量集来说并不是成本最低的选择。
D. DynamoDB 没有本机矢量相似性搜索。
Embeddings must live ALONGSIDE existing relational records in the SAME managed PostgreSQL database for transactional consistency, with minimal new technology. Which option is BEST?
嵌入必须与同一托管 PostgreSQL 数据库中的现有关系记录一起存在，以实现事务一致性，并使用最少的新技术。哪个选项最好？
- A
Amazon Aurora PostgreSQL with the pgvector extension.
带有 pgvector 扩展的 Amazon Aurora PostgreSQL。
- B
Amazon OpenSearch Serverless.
亚马逊 OpenSearch 无服务器。
- C
Amazon S3 Vectors.
亚马逊 S3 向量。
- D
Amazon Neptune Analytics.
亚马逊海王星分析。
解析 / Explanation
正确答案 / Answer: A
A. Aurora PostgreSQL with pgvector keeps vectors next to relational data in one transactional database with familiar tooling.
B. OpenSearch Serverless is excellent for vectors but is a separate system with no transactional co-location with the relational data.
C. S3 Vectors is a separate, non-transactional store and cannot co-locate with relational records.
D. Neptune is a graph store, not a relational+vector database.
A. 带有 pgvector 的 Aurora PostgreSQL 使用熟悉的工具将向量与关系数据一起保存在一个事务数据库中。
B. OpenSearch Serverless 非常适合向量，但它是一个独立的系统，不与关系数据进行事务共置。
C. S3 Vectors 是一个独立的非事务性存储，不能与关系记录共置。
D. Neptune 是一个图形存储，而不是关系+向量数据库。
Answering a question requires MULTI-HOP reasoning over relationships among connected entities, not just similarity to text. Which knowledge base option is BEST?
回答问题需要对连接实体之间的关系进行多跳推理，而不仅仅是与文本的相似性。哪个知识库选项最好？
- A
A knowledge base using GraphRAG with Amazon Neptune Analytics.
使用 GraphRAG 和 Amazon Neptune Analytics 的知识库。
- B
A vector-only knowledge base with more results returned.
返回更多结果的纯向量知识库。
- C
Aurora PostgreSQL with pgvector.
Aurora PostgreSQL 与 pgvector。
- D
A structured-data (text-to-SQL) knowledge base.
结构化数据（文本到 SQL）知识库。
解析 / Explanation
正确答案 / Answer: A
A. GraphRAG builds an entity-relationship graph that supports multi-hop reasoning over connections.
B. Returning more similar chunks does not provide explicit multi-hop relationship traversal.
C. pgvector does similarity search, not graph multi-hop reasoning.
D. Structured-data retrieval handles tabular aggregation, not entity-to-entity multi-hop reasoning.
A. GraphRAG 构建一个实体关系图，支持通过连接进行多跳推理。
B. 返回更多相似的块不提供显式的多跳关系遍历。
C. pgvector 进行相似性搜索，而不是图多跳推理。
D. 结构化数据检索处理表格聚合，而不是实体到实体的多跳推理。
Business users ask natural-language questions that require SUMS and aggregations over an Amazon Redshift warehouse, and the assistant should generate and run the queries. Which option is BEST?
业务用户提出需要通过 Amazon Redshift 仓库进行求和和聚合的自然语言问题，助手应生成并运行查询。哪个选项最好？
- A
Amazon Bedrock Knowledge Bases with structured data retrieval (text-to-SQL).
具有结构化数据检索（文本到 SQL）的 Amazon Bedrock 知识库。
- B
Embed each table row as a vector and use similarity search.
将每个表行嵌入为向量并使用相似性搜索。
- C
A GraphRAG knowledge base on Neptune Analytics.
基于 Neptune Analytics 的 GraphRAG 知识库。
- D
Amazon QuickSight Q.
亚马逊 QuickSight Q。
解析 / Explanation
正确答案 / Answer: A
A. Structured-data retrieval translates natural language into SQL against the warehouse, returning accurate aggregations.
B. Vector similarity over rows cannot compute sums or precise aggregations.
C. GraphRAG reasons over relationships, not tabular aggregation.
D. QuickSight Q answers over prepared BI datasets in a dashboard; it is not the Bedrock RAG generation path the assistant needs.
A. 结构化数据检索将自然语言转换为针对仓库的 SQL，返回准确的聚合。
B. 行上的向量相似度无法计算总和或精确聚合。
C. GraphRAG 对关系进行推理，而不是表格聚合。
D. QuickSight Q 在仪表板中回答准备好的 BI 数据集；这不是助手需要的 Bedrock RAG 生成路径。
A developer must fetch the top relevant chunks and then run them through their OWN prompt and a model NOT wired into the knowledge base. Which API is correct?
开发人员必须获取最相关的块，然后通过自己的提示符和未连接到知识库的模型来运行它们。哪个 API 是正确的？
- A
The Retrieve API.
检索 API。
- B
The RetrieveAndGenerate API.
检索并生成 API。
- C
The RetrieveAndGenerateStream API.
RetrieveAndGenerateStream API。
- D
The StartIngestionJob API.
开始摄取作业 API。
解析 / Explanation
正确答案 / Answer: A
A. Retrieve returns ranked chunks only, leaving the prompt and generation model fully under the developer's control.
B. RetrieveAndGenerate performs the generation for you, removing the custom prompt/model control required here.
C. This also performs generation (streamed), so it does not give custom prompt/model control.
D. StartIngestionJob ingests data; it does not retrieve chunks for a query.
A. 检索仅返回排序的块，将提示和生成模型完全置于开发人员的控制之下。
B. RetrieveAndGenerate 为您执行生成操作，删除此处所需的自定义提示/模型控件。
C. 这也执行生成（流式传输），因此它不提供自定义提示/模型控制。
D. StartIngestionJob摄取数据；它不会检索查询的块。
An app must show end users the exact source passages behind each answer with the LEAST custom development. Which approach is BEST?
应用程序必须以最少的定制开发向最终用户显示每个答案背后的确切源段落。哪种方法最好？
- A
Use RetrieveAndGenerate, which returns the answer with citations to the source chunks.
使用 RetrieveAndGenerate，它返回带有源块引用的答案。
- B
Use Retrieve, then build custom generation and citation-mapping code.
使用 Retrieve，然后构建自定义生成和引文映射代码。
- C
Enable model invocation logging and surface the logs.
启用模型调用日志记录并显示日志。
- D
Ask the model in the prompt to list its sources.
要求提示中的模型列出其来源。
解析 / Explanation
正确答案 / Answer: A
A. RetrieveAndGenerate produces a grounded answer plus citations with no custom attribution code.
B. This works but adds the custom generation and citation code the requirement wants to avoid.
C. Invocation logs are operator data, not user-facing citations.
D. Self-reported sources are unreliable and not verifiable citations.
A. RetrieveAndGenerate 生成有根据的答案以及引用，无需自定义归因代码。
B. 这可行，但添加了需求想要避免的自定义生成和引用代码。
C. 调用日志是操作员数据，而不是面向用户的引用。
D. 自我报告的来源是不可靠且不可验证的引用。
A multi-tenant SaaS uses one shared knowledge base and must GUARANTEE that a tenant can never retrieve another tenant's documents. Which design is the most secure and reliable?
多租户 SaaS 使用一个共享知识库，并且必须保证租户永远无法检索另一个租户的文档。哪种设计最安全可靠？
- A
Tag documents with tenant ID and apply a metadata filter derived from the authenticated tenant on every Retrieve call (or use a separate knowledge base per tenant).
使用租户 ID 标记文档，并在每次检索调用时应用从经过身份验证的租户派生的元数据过滤器（或为每个租户使用单独的知识库）。
- B
Pass the tenant filter from the client and trust it.
从客户端传递租户过滤器并信任它。
- C
Instruct the model in the system prompt to stay within the tenant.
在系统提示中指示模型留在租户内。
- D
Rely on relevance ranking to keep tenants separate.
依靠相关性排名来区分租户。
解析 / Explanation
正确答案 / Answer: A
A. A server-side filter bound to the verified identity (or full per-tenant isolation) reliably prevents cross-tenant retrieval.
B. A client-supplied, untrusted filter can be tampered with to reach other tenants' data.
C. Prompt instructions do not control retrieval and can be bypassed.
D. Ranking can still surface another tenant's documents; it is not an isolation control.
A. 绑定到已验证身份的服务器端过滤器（或完全每租户隔离）可靠地防止跨租户检索。
B. 客户端提供的不受信任的过滤器可以被篡改以获取其他租户的数据。
C. 提示指令不控制检索并且可以被绕过。
D. 排名仍然可以显示其他租户的文件；它不是隔离控制。
Relevant passages ARE being retrieved but appear LOW in the result list, so answers miss them. The team wants to fix the ordering with the least overhead. Which option is BEST?
正在检索相关段落，但在结果列表中显示为低，因此答案会错过它们。该团队希望以最少的开销来修复排序问题。哪个选项最好？
- A
Add a Bedrock reranker model to reorder the retrieved candidates by query relevance.
添加 Bedrock 重新排序器模型，以按查询相关性对检索到的候选者重新排序。
- B
Switch to hybrid search.
切换到混合搜索。
- C
Increase the number of results returned.
增加返回结果的数量。
- D
Re-ingest with semantic chunking.
通过语义分块重新摄取。
解析 / Explanation
正确答案 / Answer: A
A. A reranker reorders already-retrieved candidates so the most relevant passages move to the top.
B. Hybrid search improves whether relevant items are retrieved (recall); the problem here is ordering of items already retrieved.
C. Returning more results does not move the relevant ones to the top.
D. Re-chunking is a heavier ingestion change and does not directly reorder results.
A. 重新排序器会对已检索的候选内容重新排序，以便最相关的段落移至顶部。
B. 混合搜索提高了相关项目是否被检索（召回）；这里的问题是已经检索到的项目的排序。
C. 返回更多结果不会将相关结果移至顶部。
D. 重新分块是一种较重的摄取更改，不会直接对结果进行重新排序。
Queries that contain exact product codes and acronyms are frequently MISSED by pure semantic search. Which change BEST fixes retrieval?
纯语义搜索经常会错过包含确切产品代码和首字母缩略词的查询。哪项更改最能修复检索问题？
- A
Enable hybrid search that combines semantic vectors with keyword matching.
启用将语义向量与关键字匹配相结合的混合搜索。
- B
Add a reranker model.
添加重新排序模型。
- C
Increase the embedding dimensions.
增加嵌入尺寸。
- D
Raise the generation temperature.
提高生成温度。
解析 / Explanation
正确答案 / Answer: A
A. Hybrid search adds keyword matching so exact tokens like product codes are found alongside semantic matches.
B. A reranker reorders retrieved items but cannot surface an exact-keyword match that the vector search failed to retrieve.
C. Higher dimensions do not reliably solve exact-token matching.
D. Temperature affects generation, not retrieval of exact terms.
A. 混合搜索添加了关键字匹配，因此可以在语义匹配的同时找到产品代码等精确标记。
B. 重新排序器对检索到的项目进行重新排序，但无法显示矢量搜索未能检索到的精确关键字匹配。
C. 更高的维度不能可靠地解决精确令牌匹配问题。
D. 温度影响精确项的生成，而不是检索。
Content must be ingested into one knowledge base from Amazon S3, Confluence, and SharePoint with the LEAST custom code. What should the team do?
必须使用最少的自定义代码将内容从 Amazon S3、Confluence 和 SharePoint 提取到一个知识库中。团队应该做什么？
- A
Use the built-in Bedrock knowledge base data source connectors for each source.
对每个源使用内置的 Bedrock 知识库数据源连接器。
- B
Write custom crawlers that copy everything to S3 first.
编写自定义爬虫，首先将所有内容复制到 S3。
- C
Manually export each source to S3 on a schedule.
按计划手动将每个源导出到 S3。
- D
Fine-tune the model on each source instead of ingesting.
在每个源上微调模型而不是摄取。
解析 / Explanation
正确答案 / Answer: A
A. Managed connectors for S3, Confluence, SharePoint, and others ingest the content without custom code.
B. Custom crawlers add development and maintenance the connectors remove.
C. Manual exports are brittle and lag behind source changes.
D. Fine-tuning is not ingestion and does not keep content current.
A. S3、Confluence、SharePoint 等的托管连接器无需自定义代码即可摄取内容。
B. 自定义爬虫添加了连接器删除的开发和维护。
C. 手动导出很脆弱并且落后于源更改。
D. 微调不是摄取，也不能使内容保持最新状态。
A RetrieveAndGenerate chat must keep context across follow-up questions with the LEAST custom state management. Which approach is BEST?
RetrieveAndGenerate 聊天必须使用最少的自定义状态管理来保留后续问题的上下文。哪种方法最好？
- A
Pass a session identifier to RetrieveAndGenerate so Bedrock maintains the conversation context.
将会话标识符传递给 RetrieveAndGenerate，以便 Bedrock 维护对话上下文。
- B
Store every message in DynamoDB and rebuild the prompt each turn.
将每条消息存储在 DynamoDB 中并每轮重建提示。
- C
Re-embed the chat history into the vector store each turn.
每回合将聊天历史记录重新嵌入矢量存储中。
- D
Use a model with a larger context window.
使用具有较大上下文窗口的模型。
解析 / Explanation
正确答案 / Answer: A
A. The session ID preserves multi-turn context without the developer building custom conversation state.
B. Manual state assembly is more work than the built-in session support.
C. Re-embedding history is costly and is not how session context is maintained.
D. A larger window does not automatically carry prior turns into each call.
A. 会话 ID 保留多轮上下文，而无需开发人员构建自定义对话状态。
B. 手动状态组装比内置会话支持需要更多工作。
C. 重新嵌入历史记录的成本很高，并且不是维护会话上下文的方式。
D. 较大的窗口不会自动将优先轮流带入每次呼叫。
Complex, multi-part questions retrieve passages for only ONE part, yielding incomplete answers. Which built-in capability BEST addresses this?
复杂的多部分问题仅检索其中一个部分的段落，从而产生不完整的答案。哪种内置功能最能解决这个问题？
- A
Enable query decomposition (reformulation) so the knowledge base splits the question into sub-queries.
启用查询分解（重新表述），以便知识库将问题拆分为子查询。
- B
Increase the number of results returned.
增加返回结果的数量。
- C
Add a reranker model.
添加重新排序模型。
- D
Switch to a larger generation model.
切换到更大一代的型号。
解析 / Explanation
正确答案 / Answer: A
A. Decomposition retrieves passages for each sub-question, improving coverage and completeness.
B. More results add noise but do not ensure each sub-question is covered.
C. Reranking reorders retrieved items but cannot fetch passages for an unaddressed sub-question.
D. A larger model cannot use passages that were never retrieved.
A. 分解检索每个子问题的段落，提高覆盖范围和完整性。
B. 更多结果会增加噪音，但不能确保涵盖每个子问题。
C. 重新排序会对检索到的项目重新排序，但无法获取未解决的子问题的段落。
D. 较大的模型不能使用从未检索到的段落。
New and deleted documents in Amazon S3 must be reflected in a Bedrock knowledge base quickly, WITHOUT re-scanning the entire source each time, in a scalable and resilient way. Which design is BEST?
Amazon S3 中的新文档和已删除文档必须以可扩展且有弹性的方式快速反映在 Bedrock 知识库中，而无需每次重新扫描整个源。哪种设计最好？
- A
Send S3 object-created/deleted events to an SQS queue, and have a Lambda call the incremental IngestKnowledgeBaseDocuments / DeleteKnowledgeBaseDocuments APIs.
将 S3 对象创建/删除事件发送到 SQS 队列，并让 Lambda 调用增量 IngestKnowledgeBaseDocuments / DeleteKnowledgeBaseDocuments API。
- B
On every change, call StartIngestionJob to sync the whole data source.
每次更改时，调用 StartIngestionJob 来同步整个数据源。
- C
Invoke a Lambda directly from S3 events with no queue.
直接从 S3 事件调用 Lambda，无需队列。
- D
Run an EventBridge Scheduler every 5 minutes to diff the bucket.
每 5 分钟运行一次 EventBridge Scheduler 来比较存储桶。
解析 / Explanation
正确答案 / Answer: A
A. Event-driven SQS buffering with the incremental document APIs updates only changed items and adds retries/resilience.
B. StartIngestionJob re-scans the source, which is inefficient and not incremental.
C. Direct invocation lacks buffering and retries and can fail under load.
D. Polling on a timer is not near-real-time and is less efficient than event-driven updates.
A. 事件驱动的 SQS 缓冲与增量文档 API 仅更新更改的项目并增加重试/弹性。
B. StartIngestionJob重新扫描源，效率低且不增量。
C. 直接调用缺乏缓冲和重试，并且可能在负载下失败。
D. 计时器轮询不是近乎实时的，并且效率低于事件驱动的更新。
A team wants grounded answers WITH citations over their documents, with the least operational overhead and no retrieval pipeline to build. Which option is BEST?
团队希望得到有依据的答案，并对其文档进行引用，并且运营开销最少，并且无需构建检索管道。哪个选项最好？
- A
Amazon Bedrock Knowledge Bases (managed RAG).
Amazon Bedrock 知识库（托管 RAG）。
- B
Amazon Kendra by itself.
亚马逊 Kendra 本身。
- C
A custom pipeline with Lambda and a self-managed vector database.
使用 Lambda 和自我管理向量数据库的自定义管道。
- D
Fine-tune a model on the documents.
在文档上微调模型。
解析 / Explanation
正确答案 / Answer: A
A. Knowledge Bases manage chunking, embeddings, the vector store, retrieval, and citations as a service.
B. Kendra is enterprise search; on its own it does not provide the managed FM generation-with-citations RAG flow.
C. A hand-built pipeline maximizes operational overhead.
D. Fine-tuning is not RAG and does not provide citations or freshness.
A. 知识库将分块、嵌入、向量存储、检索和引用作为服务进行管理。
B. Kendra 是企业搜索；它本身不提供托管 FM 生成与引文 RAG 流程。
C. 手工构建的管道可以最大限度地提高运营开销。
D. 微调不是 RAG，也不提供引用或新鲜度。
A RAG assistant returns confident answers that the retrieved passages do not actually support. The team wants a MANAGED, tunable way to block ungrounded responses. Which option is BEST?
RAG 助手返回了检索到的段落实际上并不支持的自信答案。该团队需要一种可管理的、可调整的方式来阻止不切实际的响应。哪个选项最好？
- A
Enable the Amazon Bedrock Guardrails contextual grounding check with a tuned threshold.
使用调整后的阈值启用 Amazon Bedrock Guardrails 上下文接地检查。
- B
Add a reranker model.
添加重新排序模型。
- C
Lower the temperature.
降温。
- D
Switch to a larger model.
切换到更大的型号。
解析 / Explanation
正确答案 / Answer: A
A. The contextual grounding check scores whether a response is supported by the source and can block low-scoring answers.
B. A reranker improves ordering of retrieved passages but does not verify or block ungrounded generation.
C. Lower randomness does not guarantee grounding in the retrieved context.
D. A larger model can still produce ungrounded answers without a grounding control.
A. 上下文基础检查对响应是否得到来源支持进行评分，并可以阻止低分答案。
B. 重新排序器可以改善检索到的段落的排序，但不会验证或阻止未接地的生成。
C. 较低的随机性并不能保证检索到的上下文中的基础。
D. 较大的模型仍然可以在没有接地控制的情况下产生不接地的答案。
Retrieval must MATCH on small, precise text units but then give the language model the BROADER surrounding context for those matches. Which chunking strategy is BEST?
检索必须匹配小而精确的文本单元，然后为语言模型提供这些匹配的更广泛的周围上下文。哪种分块策略最好？
- A
Hierarchical (parent-child) chunking.
分层（父子）分块。
- B
Semantic chunking.
语义分块。
- C
Large fixed-size chunking.
大型固定大小分块。
- D
No chunking.
没有分块。
解析 / Explanation
正确答案 / Answer: A
A. Hierarchical chunking matches on small child chunks for precision and supplies the larger parent chunk as context for generation.
B. Semantic chunking yields coherent chunks but does not provide the parent-context-on-child-match behavior required here.
C. Large fixed chunks sacrifice precise matching on small units.
D. Treating each document as one chunk loses the precise small-unit matching.
A. 分层分块匹配较小的子块以提高精度，并提供较大的父块作为生成的上下文。
B. 语义分块产生连贯的块，但不提供此处所需的父上下文子匹配行为。
C. 大的固定块牺牲了小单元的精确匹配。
D. 将每个文档视为一个块会失去精确的小单元匹配。