AWS AIP-C01 开源练习 · 困难 · 第 01 套

A. For steady, high, predictable volume, a committed Provisioned Throughput model unit lowers cost per token and guarantees dedicated capacity.

B. On-demand pricing suits spiky or unpredictable low volume; at steady high volume it is not the lowest cost and does not reserve capacity.

C. Batch inference is for offline, latency-tolerant jobs; it cannot serve steady real-time traffic.

D. Larger outputs raise cost per call and do not reduce the number of distinct summaries needed.

A. 为了获得稳定、高、可预测的交易量，承诺的预配置吞吐量模型单位可以降低每个代币的成本并保证专用容量。

B. 按需定价适合尖峰或不可预测的小批量；在稳定的高产量下，它不是最低的成本，并且不保留产能。

C. 批量推理适用于离线、耐延迟的作业；它无法提供稳定的实时流量。

D. 较大的输出会增加每次调用的成本，并且不会减少所需的不同摘要的数量。

A. Guardrails provide configurable, model-agnostic policies for denied topics, content filters, word filters, and PII redaction that apply consistently across FMs.

B. IAM controls who can call the model, not what content the model may produce or must redact.

C. Alarms detect operational anomalies after the fact; they do not filter or redact content.

D. Temperature affects randomness, not topic blocking, content filtering, or PII redaction.

A. Guardrails 为被拒绝的主题、内容过滤器、单词过滤器和 PII 编辑提供可配置的、与模型无关的策略，这些策略在 FM 中一致应用。

B. IAM 控制谁可以调用模型，而不控制模型可能生成或必须编辑的内容。

C. 警报在事后检测到操作异常；他们不会过滤或编辑内容。

D. 温度影响随机性，而不影响主题阻止、内容过滤或 PII 编辑。

A. Right-sizing to a smaller/distilled model reduces inference time, and response streaming lowers perceived latency by showing tokens as they are generated.

B. The largest model maximizes capability but adds latency and cost that routine questions do not require.

C. Batch inference is asynchronous and offline; it is unsuitable for interactive latency.

D. Larger outputs increase generation time, worsening perceived latency.

A. 调整为较小/精炼模型可以减少推理时间，响应流通过显示生成的令牌来降低感知延迟。

B. 最大的模型可以最大化功能，但会增加常规问题不需要的延迟和成本。

C. 批量推理是异步且离线的；它不适合交互延迟。

D. 较大的输出会增加生成时间，从而加剧感知延迟。

A. Operational Excellence emphasizes automating evaluation and observability so quality is measured continuously and improvements are driven by data.

B. A one-time check does not catch regressions introduced by later model or prompt changes.

C. Waiting for complaints is reactive and lets defects reach users undetected.

D. Infrastructure health says nothing about generation quality or drift.

A. 卓越运营强调自动化评估和可观察性，以便持续衡量质量并由数据驱动改进。

B. 一次性检查不会捕获后续模型或提示更改引入的回归。

C. 等待投诉是被动的，会让缺陷在未被发现的情况下到达用户手中。

D. 基础设施的健康状况与发电质量或漂移无关。

A. Least-privilege identity-based policies with condition keys restrict who can call models and can enforce that guardrails are applied.

B. Customer-managed encryption and private connectivity protect data confidentiality and reduce exposure of prompts and outputs.

C. Long-lived embedded keys are a major security risk and violate credential best practices.

D. Unencrypted transport exposes sensitive prompt and response data.

E. Disabling logging removes the audit trail needed for security monitoring and compliance.

A. 带有条件键的基于身份的最小权限策略限制了谁可以调用模型，并可以强制应用护栏。

B. 客户管理的加密和专用连接可保护数据机密性并减少提示和输出的暴露。

C. 长期嵌入的密钥是一个主要的安全风险，并且违反了凭证最佳实践。

D. 未加密的传输会暴露敏感的提示和响应数据。

E. 禁用日志记录会删除安全监控和合规性所需的审核跟踪。

A. Batch inference processes large datasets asynchronously at lower cost, and a right-sized model avoids over-provisioning compute for a latency-tolerant job.

B. A continuously provisioned endpoint wastes capacity for a once-daily, latency-tolerant job.

C. High-volume real-time invocation is more expensive and unnecessary when results are needed only the next day.

D. This maximizes cost and energy use with no benefit for a non-interactive workload.

A. 批量推理以较低的成本异步处理大型数据集，并且大小合适的模型可以避免为容忍延迟的作业过度配置计算。

B. 连续配置的端点会浪费每日一次的延迟容忍作业的容量。

C. 当仅在第二天需要结果时，大容量实时调用成本更高且不必要。

D. 这最大限度地提高了成本和能源使用，但对非交互式工作负载没有任何好处。

A. Prompt caching stores and reuses repeated prompt prefixes, cutting both the tokens billed and the latency for the unchanging context.

B. Temperature changes output randomness and has no effect on cost or latency of repeated context.

C. Provisioned Throughput reserves capacity but still reprocesses the repeated context on every call and changes the pricing model significantly.

D. A larger context window does not reduce the cost of reprocessing the same tokens each request.

A. 提示缓存存储并重用重复的提示前缀，从而减少计费令牌和不变上下文的延迟。

B. 温度会改变输出的随机性，并且对重复上下文的成本或延迟没有影响。

C. 预配置吞吐量保留了容量，但仍重新处理每次调用的重复上下文，并显着改变了定价模型。

D. 较大的上下文窗口不会降低每个请求重新处理相同令牌的成本。

A. Ground Truth provides managed data-labeling workflows, workforce options (private, vendor, or Mechanical Turk), and quality features such as annotation consolidation.

B. DataBrew cleans and normalizes tabular data; it does not run human annotation workflows.

C. Comprehend performs automated NLP such as entity and PII detection, not managed human labeling.

D. Macie discovers and classifies sensitive data; it does not label datasets for training.

A. Ground Truth 提供托管数据标签工作流程、劳动力选项（私人、供应商或 Mechanical Turk）以及注释整合等质量功能。

B. DataBrew 清理并规范化表格数据；它不运行人工注释工作流程。

C. Comprehend 执行自动化 NLP，例如实体和 PII 检测，而不是托管人工标记。

D. Macie 发现敏感数据并对其进行分类；它不会标记用于训练的数据集。

A. Data Wrangler offers a visual, low-code interface for importing, cleaning, transforming, and exporting tabular data for ML, reducing custom preprocessing code.

B. Athena queries data in S3 with SQL but is not a guided data-preparation and feature-engineering tool.

C. QuickSight is for business-intelligence dashboards, not ML data preparation.

D. Kendra is an enterprise search service, not a data-prep tool.

A. Data Wrangler 提供了一个可视化的低代码界面，用于导入、清理、转换和导出 ML 的表格数据，从而减少自定义预处理代码。

B. Athena 使用 SQL 查询 S3 中的数据，但不是引导式数据准备和特征工程工具。

C. QuickSight 适用于商业智能仪表板，而不是 ML 数据准备。

D. Kendra 是一种企业搜索服务，而不是数据准备工具。

A. Glue runs serverless Apache Spark ETL that cleans, transforms, and partitions large datasets without provisioning or managing clusters.

B. EMR is powerful but requires provisioning and managing cluster capacity, adding operational overhead.

C. Self-managed EC2 batch processing maximizes operational overhead.

D. Per-file Lambda parsing is hard to coordinate at scale and hits timeout and orchestration limits for large ETL.

A. Glue 运行无服务器 Apache Spark ETL，可以清理、转换和分区大型数据集，而无需配置或管理集群。

B. EMR 功能强大，但需要配置和管理集群容量，从而增加了运营开销。

C. 自我管理的 EC2 批处理可最大限度地提高运营开销。

D. 每个文件的 Lambda 解析很难大规模协调，并且会遇到大型 ETL 的超时和编排限制。

A. Macie uses managed data identifiers and machine learning to discover and classify sensitive data and PII across S3 at scale.

B. Comprehend detects and can redact PII within individual text inputs but is not designed to inventory sensitive data across many buckets.

C. Inspector assesses workloads for software vulnerabilities, not sensitive-data classification.

D. GuardDuty is a threat-detection service for accounts and workloads, not data classification.

A. Macie 使用托管数据标识符和机器学习来大规模发现和分类 S3 中的敏感数据和 PII。

B. Comprehend 可以检测并编辑各个文本输入中的 PII，但其设计目的不是跨多个存储桶清点敏感数据。

C. Inspector 评估软件漏洞的工作负载，而不是敏感数据分类。

D. GuardDuty 是针对帐户和工作负载的威胁检测服务，而不是数据分类服务。

A. Comprehend can detect PII entities in text and redact or mask them, producing sanitized transcripts suitable for training.

B. Macie classifies and reports on sensitive data in S3 but does not redact PII inside text for downstream training.

C. Textract extracts text from scanned documents; it does not specialize in PII redaction of conversational text.

D. Rekognition analyzes images and video, not text PII.

A. Comprehend 可以检测文本中的 PII 实体并对其进行编辑或屏蔽，从而生成适合训练的经过净化的转录本。

B. Macie 对 S3 中的敏感数据进行分类和报告，但不会编辑文本内的 PII 以进行下游训练。

C. Textract 从扫描文档中提取文本；它并不专注于会话文本的 PII 编辑。

D. Rekognition 分析图像和视频，而不是文本 PII。

A. Clarify computes pre-training bias metrics on datasets (and post-training bias and feature-attribution explanations on models).

B. Model Monitor detects drift on a deployed endpoint, not bias in a training dataset before training.

C. CloudWatch monitors operational metrics, not dataset bias.

D. Comprehend performs NLP tasks, not dataset bias measurement.

A. Clarify 计算数据集的训练前偏差指标（以及模型的训练后偏差和特征归因解释）。

B. 模型监视器检测已部署端点上的漂移，而不是训练前训练数据集中的偏差。

C. CloudWatch 监控运营指标，而不是数据集偏差。

D. Comprehend 执行 NLP 任务，而不是数据集偏差测量。

A. Bedrock customization expects labeled examples as JSONL records in S3 that match the target model's schema.

B. An unstructured PDF is not a labeled training dataset; that pattern suits RAG ingestion, not supervised fine-tuning.

C. Unlabeled text is for continued pre-training, not supervised fine-tuning, which needs labeled pairs.

D. Fine-tuning reads training data from S3 in the required file format, not from DynamoDB.

A. Bedrock 自定义期望标记示例为 S3 中与目标模型架构匹配的 JSONL 记录。

B. 非结构化 PDF 不是带标签的训练数据集；该模式适合 RAG 摄取，而不是监督微调。

C. 未标记的文本用于持续的预训练，而不是有监督的微调，这需要标记对。

D. 微调以所需的文件格式从 S3 读取训练数据，而不是从 DynamoDB 读取。

A. Model Cards capture intended use, training details, evaluation, and risk information in a standardized, versioned, auditable document.

B. CloudTrail records API activity, not model documentation and governance metadata.

C. Config tracks resource configuration compliance, not model documentation.

D. A custom README lacks the standardized structure, versioning, and governance integration of Model Cards.

A. 模型卡在标准化、版本化、可审核的文档中捕获预期用途、培训详细信息、评估和风险信息。

B. CloudTrail 记录 API 活动，而不是模型文档和治理元数据。

C. Config 跟踪资源配置合规性，而不是模型文档。

D. 自定义自述文件缺乏模型卡的标准化结构、版本控制和治理集成。

A. Removing duplicates and normalizing inconsistent text reduces bias toward repeated samples and improves the signal the model learns from.

B. A non-overlapping evaluation split prevents train/test leakage and gives an honest measure of generalization.

C. Duplication inflates size artificially and encourages overfitting to repeated examples.

D. Including PII creates privacy and compliance risk and does not improve task quality.

E. Blindly adding epochs is a training hyperparameter choice, not data preparation, and risks overfitting.

A. 删除重复项和标准化不一致的文本可以减少对重复样本的偏差，并改善模型学习的信号。

B. 不重叠的评估分割可以防止训练/测试泄漏，并提供诚实的泛化度量。

C. 重复会人为地夸大大小，并导致对重复示例的过度拟合。

D. 包含 PII 会带来隐私和合规风险，并且不会提高任务质量。

E. 盲目添加纪元是训练超参数的选择，而不是数据准备，并且存在过度拟合的风险。

A. Nova Micro is the text-only, lowest-latency, lowest-cost Nova model, well suited to high-volume, simple text tasks such as intent classification.

B. Nova Pro is the most capable multimodal Nova model, with higher cost and latency than a simple classification task requires.

C. Nova Canvas generates images and is not used for text classification.

D. Nova Reel generates video and is irrelevant to text intent classification.

A. Nova Micro 是纯文本、延迟最低、成本最低的 Nova 模型，非常适合大容量、简单的文本任务，例如意图分类。

B. Nova Pro 是功能最强大的多模式 Nova 模型，与简单的分类任务相比，其成本和延迟更高。

C. Nova Canvas 生成图像，不用于文本分类。

D. Nova Reel 生成视频，与文本意图分类无关。

A. Bedrock model distillation generates data from a teacher model to fine-tune a smaller student that approximates its quality at lower cost and latency.

B. Cross-Region inference distributes load across Regions for throughput; it does not reduce per-call model size or cost.

C. Provisioned Throughput reserves capacity for a chosen model but does not create a smaller, cheaper model.

D. Prompt caching reuses repeated prompt prefixes but does not change the model used for inference.

A. 基岩模型蒸馏从教师模型中生成数据，以对较小的学生进行微调，以较低的成本和延迟来近似其质量。

B. 跨区域推理将负载分布到各个区域以提高吞吐量；它不会减少每次调用模型的大小或成本。

C. 预配置吞吐量为所选模型保留容量，但不会创建更小、更便宜的模型。

D. 提示缓存重用重复的提示前缀，但不会更改用于推理的模型。

A. RAG retrieves the current documents at query time, so updated content is reflected immediately without retraining the model.

B. Repeated fine-tuning is costly and operationally heavy and still lags behind document changes between runs.

C. Continued pre-training is even heavier than fine-tuning and is not how you keep answers current.

D. Stuffing all documents into the prompt is costly, hits context limits, and does not scale as the corpus grows.

A. RAG 在查询时检索当前文档，因此更新的内容会立即反映，而无需重新训练模型。

B. 重复的微调成本高昂，操作繁重，并且仍然落后于运行之间的文档更改。

C. 持续的预训练比微调更重要，并且不是保持答案最新的方式。

D. 将所有文档填充到提示中成本高昂，会达到上下文限制，并且不会随着语料库的增长而扩展。

A. Fine-tuning teaches consistent style and formatting from examples and is appropriate when the behavior, not freshness of facts, is the gap.

B. RAG supplies fresh facts but does not reliably enforce tone and output formatting.

C. Higher temperature increases variability, working against a consistent voice and format.

D. Provisioned Throughput affects capacity and cost, not the model's style or formatting behavior.

A. 微调可以从示例中教授一致的风格和格式，并且当行为（而不是事实的新鲜度）存在差距时是适当的。

B. RAG 提供了新的事实，但不能可靠地强制执行语气和输出格式。

C. 较高的温度会增加可变性，不利于一致的语音和格式。

D. 预置吞吐量影响容量和成本，而不是模型的样式或格式化行为。

A. Continued pre-training adapts a model to domain language using large volumes of unlabeled text.

B. Fine-tuning needs labeled examples and targets specific task behavior rather than broad domain-language adaptation.

C. RAG injects retrieved context at inference time but does not change what the model has learned.

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. 微调需要标记的示例并针对特定的任务行为，而不是广泛的领域语言适应。

C. RAG 在推理时注入检索到的上下文，但不会更改模型学到的内容。

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。

A. A low temperature reduces randomness, making outputs more deterministic and repeatable.

B. A lower top-p narrows the token pool to the most probable tokens, reducing variability.

C. Higher temperature increases randomness, the opposite of the requirement.

D. A large top-k widens the candidate pool and increases variability.

E. Max tokens controls output length, not determinism.

A. 低温会降低随机性，使输出更具确定性和可重复性。

B. 较低的 top-p 将代币池缩小到最可能的代币，从而减少可变性。

C. 较高的温度会增加随机性，这与要求相反。

D. 较大的 top-k 扩大了候选池并增加了可变性。

E. 最大令牌控制输出长度，而不是确定性。

A. Batch inference processes large datasets asynchronously at a lower price point, ideal for latency-tolerant, high-volume jobs.

B. Per-record real-time calls are more expensive and unnecessary for a weekly, offline report.

C. Reserved capacity running 24/7 wastes money for a once-weekly job.

D. Max tokens does not batch records and would not reduce overall cost for this workload.

A. 批量推理以较低的价格异步处理大型数据集，非常适合延迟容忍的大批量作业。

B. 对于每周的离线报告来说，每条记录的实时调用成本更高，而且没有必要。

C. 保留容量运行 24/7 每周一次的工作会浪费金钱。

D. 最大令牌不会批量记录，也不会降低此工作负载的总体成本。

A. On-demand pricing charges per request with no commitment, which is most cost-effective for spiky, unpredictable, low-utilization traffic.

B. A commitment pays for reserved capacity that sits idle during the long quiet periods.

C. Batch is offline and cannot serve interactive, bursty requests.

D. An always-on endpoint incurs cost during idle periods, defeating the goal.

A. 按请求按需定价，无需承诺，这对于尖峰、不可预测、低利用率的流量来说最具成本效益。

B. 承诺为在长时间安静期间闲置的保留容量付费。

C. 批处理处于脱机状态，无法处理交互式突发请求。

D. 始终在线的端点会在空闲期间产生成本，从而达不到目标。

A. Bedrock model evaluation supports automatic LLM-as-a-judge scoring across quality dimensions at scale with minimal setup.

B. Human-only evaluation is accurate but operationally heavy and slow at scale.

C. Ad hoc manual review is not systematic or scalable for a reliable comparison.

D. Latency metrics do not measure summarization relevance or coherence.

A. 基岩模型评估支持以最少的设置在质量维度上以法学硕士为评委进行自动评分。

B. 仅人工评估是准确的，但操作繁重且规模缓慢。

C. 临时手动审核不具有系统性或可扩展性，无法进行可靠的比较。

D. 延迟指标不衡量摘要的相关性或一致性。

A. Converse provides a unified, model-agnostic interface for multi-turn messages and tool use across supported Bedrock models.

B. Per-model InvokeModel payloads require model-specific request/response handling, which Converse abstracts away.

C. Building adapters reinvents what the Converse API already standardizes.

D. API Gateway mapping does not normalize differing model message and tool-use formats.

A. Converse 为跨受支持的 Bedrock 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。

B. 每个模型的 InvokeModel 有效负载需要模型特定的请求/响应处理，Converse 对此进行了抽象。

C. 构建适配器重塑了 Converse API 已经标准化的内容。

D. API 网关映射不会规范不同的模型消息和工具使用格式。

A. Latency-optimized inference reduces response latency for supported models without an application rewrite.

B. A larger model typically increases latency rather than reducing it.

C. Batch inference is asynchronous and increases, not decreases, time-to-response.

D. Retries handle failures; they do not lower per-request latency.

A. 延迟优化推理可减少支持模型的响应延迟，而无需重写应用程序。

B. 较大的模型通常会增加而不是减少延迟。

C. 批量推理是异步的，会增加而不是减少响应时间。

D. 重试处理失败；它们不会降低每个请求的延迟。

A. Custom Model Import lets you bring supported custom model weights into Bedrock and invoke them through the standard on-demand Bedrock APIs without managing servers.

B. A SageMaker endpoint can serve the model but requires managing the endpoint and does not expose it through the Bedrock APIs.

C. Running the model on ECS means operating your own inference infrastructure.

D. Provisioned Throughput reserves capacity for Bedrock models, not for importing your own weights.

A. 自定义模型导入可让您将受支持的自定义模型权重引入 Bedrock 中，并通过标准的按需 Bedrock API 调用它们，而无需管理服务器。

B. SageMaker 端点可以为模型提供服务，但需要管理端点并且不通过 Bedrock API 公开它。

C. 在 ECS 上运行模型意味着操作您自己的推理基础设施。

D. 预置吞吐量为基岩模型保留容量，而不是用于导入您自己的权重。

A. A multimodal FM can reason over image inputs and produce natural-language answers about their content.

B. Rekognition returns labels and detections but does not hold a natural-language Q&A about an image.

C. Textract extracts text from documents and forms; it does not answer visual questions about general photos.

D. Text embeddings represent text for search; they do not analyze image content or answer questions.

A. 多模态 FM 可以对图像输入进行推理并生成有关其内容的自然语言答案。

B. Rekognition 返回标签和检测结果，但不提供有关图像的自然语言问答。

C. Textract 从文档和表单中提取文本；它不回答有关一般照片的视觉问题。

D. 文本嵌入表示用于搜索的文本；他们不分析图像内容或回答问题。

A. Titan Text Embeddings V2 lets you select among multiple output dimension sizes to trade storage/cost against accuracy.

B. An image generation model does not produce text embeddings for search.

C. Claude is a text-generation model, not a dedicated embeddings model.

D. Comprehend provides NLP analyses, not configurable embedding vectors for a vector store.

A. Titan Text Embeddings V2 允许您在多个输出维度大小中进行选择，以在存储/成本和准确性之间进行权衡。

B. 图像生成模型不会生成用于搜索的文本嵌入。

C. Claude 是一个文本生成模型，而不是专用的嵌入模型。

D. Comprehend 提供 NLP 分析，而不是向量存储的可配置嵌入向量。

A. OpenSearch Serverless is a natively integrated, fully managed vector store for Bedrock knowledge bases that scales automatically for frequent low-latency queries.

B. A provisioned cluster requires capacity planning, scaling, and maintenance, adding operational overhead.

C. DynamoDB has no native vector similarity search and would require custom, slower logic.

D. Plain S3 has no vector index or similarity search capability.

A. OpenSearch Serverless 是一个原生集成的、完全托管的 Bedrock 知识库矢量存储，可自动扩展以适应频繁的低延迟查询。

B. 预配置的集群需要容量规划、扩展和维护，从而增加了运营开销。

C. DynamoDB 没有本机向量相似性搜索，需要自定义、速度较慢的逻辑。

D. Plain S3 没有向量索引或相似性搜索功能。

A. Bedrock knowledge bases support metadata attributes and filter expressions so retrieval returns only chunks matching the user's department.

B. Prompt instructions cannot reliably prevent irrelevant or unauthorized chunks from being retrieved and used.

C. Temperature has nothing to do with which documents are retrieved.

D. Truncation does not enforce department-level access and may still expose other departments' content.

A. Bedrock 知识库支持元数据属性和过滤器表达式，因此检索仅返回与用户部门匹配的块。

B. 提示指令不能可靠地防止不相关或未经授权的块被检索和使用。

C. 温度与检索哪些文档无关。

D. 截断不会强制执行部门级访问，但仍可能暴露其他部门的内容。

A. RetrieveAndGenerate produces a grounded answer and returns citations identifying the retrieved passages used, with no custom attribution code.

B. Stuffing the corpus into the prompt does not produce reliable citations and does not scale.

C. Invocation logs are for operators and do not provide user-facing source citations.

D. Asking the model to self-report sources is unreliable; max tokens does not create verifiable citations.

A. RetrieveAndGenerate 生成一个有根据的答案并返回标识所使用的检索到的段落的引文，没有自定义归因代码。

B. 将语料库塞进提示中不会产生可靠的引用，也不会扩展。

C. 调用日志适用于操作员，不提供面向用户的源引用。

D. 要求模型自我报告来源是不可靠的； max tokens 不会创建可验证的引用。

A. Retrieve returns the relevant passages only, letting the developer control the prompt and generation model themselves.

B. RetrieveAndGenerate performs the generation for you, which removes the custom prompt/model control the developer wants.

C. StartIngestionJob ingests data into the knowledge base; it does not retrieve chunks for a query.

D. That API runs fine-tuning, unrelated to retrieving chunks.

A. Retrieve仅返回相关段落，让开发者自己控制提示和生成模型。

B. RetrieveAndGenerate 为您执行生成，这会删除开发人员想要的自定义提示/模型控件。

C. StartIngestionJob将数据摄取到知识库中；它不会检索查询的块。

D. 该 API 运行微调，与检索块无关。

A. The contextual grounding (and relevance) check scores whether a response is grounded in the provided source and can block ungrounded answers.

B. Higher temperature increases variability and tends to worsen, not reduce, ungrounded output.

C. Pricing mode has no effect on grounding.

D. Removing context guarantees the answer cannot be grounded, the opposite of the goal.

A. 上下文基础（和相关性）检查对响应是否基于所提供的来源进行评分，并可以阻止不可靠的答案。

B. 较高的温度会增加变异性，并且往往会恶化而不是减少未接地的输出。

C. 定价模式对接地没有影响。

D. 删除上下文会保证答案无法扎根，这与目标相反。

A. Enforcing a server-side metadata filter bound to the verified tenant identity (or full isolation via per-tenant knowledge bases) reliably prevents cross-tenant retrieval.

B. Prompt instructions do not control retrieval and can be bypassed, so other tenants' chunks can still be retrieved.

C. A client-supplied, untrusted filter can be tampered with to access other tenants' data.

D. Relevance ranking can still surface another tenant's documents; it is not an isolation mechanism.

A. 强制将服务器端元数据过滤器绑定到已验证的租户身份（或通过每个租户知识库进行完全隔离）可以可靠地防止跨租户检索。

B. 提示指令不控制检索，可以绕过，因此仍然可以检索其他租户的块。

C. 客户端提供的不受信任的过滤器可以被篡改以访问其他租户的数据。

D. 相关性排名仍然可以显示其他租户的文件；它不是一种隔离机制。

A. GraphRAG with Neptune Analytics builds a graph of entities and relationships, enabling retrieval that reasons over connections rather than only vector similarity.

B. Pure vector similarity does not capture explicit relationships among connected entities.

C. DynamoDB stores items but provides no graph reasoning or managed RAG integration.

D. Athena runs SQL over files and is not a graph-aware retrieval system for RAG.

A. GraphRAG 与 Neptune Analytics 构建实体和关系图，支持通过连接推理而不仅仅是矢量相似性的检索。

B. 纯向量相似性不能捕获连接实体之间的显式关系。

C. DynamoDB 存储项目，但不提供图形推理或托管 RAG 集成。

D. Athena 在文件上运行 SQL，并不是 RAG 的图形感知检索系统。

A. Structured-data knowledge bases translate natural-language questions into SQL against the connected warehouse, returning accurate analytical results.

B. Vector similarity over rows cannot perform aggregations or precise analytical queries.

C. Comprehend analyzes text but does not generate SQL or query a warehouse.

D. Fine-tuning does not give live, accurate query results over changing data the way text-to-SQL retrieval does.

A. 结构化数据知识库将自然语言问题转换为针对互联仓库的 SQL，返回准确的分析结果。

B. 行上的向量相似度无法执行聚合或精确的分析查询。

C. Comprehend 分析文本，但不生成 SQL 或查询仓库。

D. 微调无法像文本到 SQL 检索那样针对变化的数据提供实时、准确的查询结果。

A. S3 Vectors offers serverless, low-cost vector storage and search at massive scale, ideal for large corpora queried infrequently with pay-for-use pricing.

B. A continuously provisioned cluster is expensive for data queried only a few times a month.

C. Keeping high-performance serverless capacity ready is more expensive than needed for rare queries.

D. RDS requires running instances and is not the lowest-cost option for rarely queried, very large vector sets.

A. S3 Vectors 提供无服务器、低成本矢量存储和大规模搜索，非常适合不经常查询且按使用付费的大型语料库。

B. 对于每月仅查询几次的数据，持续配置的集群非常昂贵。

C. 保持高性能无服务器容量的就绪性比罕见查询所需的成本更高。

D. RDS 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。

A. Aurora PostgreSQL with pgvector is a supported Bedrock vector store and lets the team keep vectors next to relational data using PostgreSQL tooling.

B. Introducing a separate self-managed system adds operational overhead the team is trying to avoid.

C. DynamoDB is not a supported vector store and lacks native similarity search.

D. In-memory storage is not durable and cannot serve as a knowledge base vector store.

A. 带有 pgvector 的 Aurora PostgreSQL 是受支持的 Bedrock 矢量存储，让团队可以使用 PostgreSQL 工具将矢量保存在关系数据旁边。

B. 引入单独的自我管理系统会增加团队试图避免的运营开销。

C. DynamoDB 不是受支持的矢量存储，并且缺乏本机相似性搜索。

D. 内存存储不持久，不能用作知识库向量存储。

A. Bedrock knowledge bases provide managed connectors for sources such as S3, Confluence, SharePoint, Salesforce, and web crawlers, avoiding custom ingestion code.

B. Custom crawlers add development and maintenance overhead that the built-in connectors remove.

C. Manual exports are brittle and lag behind source changes compared with managed connectors.

D. Fine-tuning is not how you ingest changing documents for retrieval.

A. Bedrock 知识库为 S3、Confluence、SharePoint、Salesforce 和网络爬虫等源提供托管连接器，避免自定义摄取代码。

B. 自定义爬虫增加了内置连接器消除的开发和维护开销。

C. 与托管连接器相比，手动导出很脆弱并且滞后于源更改。

D. 微调并不是获取不断变化的文档以进行检索的方式。

A. RetrieveAndGenerate supports a session ID that preserves multi-turn context without the developer building custom conversation state.

B. Re-embedding the conversation each turn is unnecessary and costly and is not how session context is maintained.

C. Manual state assembly adds development overhead compared with built-in session support.

D. A larger window does not automatically carry prior turns into each RetrieveAndGenerate call.

A. RetrieveAndGenerate 支持会话 ID，该会话 ID 可以保留多轮上下文，而无需开发人员构建自定义对话状态。

B. 每轮重新嵌入对话是不必要的且成本高昂，并且不是维护会话上下文的方式。

C. 与内置会话支持相比，手动状态组装增加了开发开销。

D. 较大的窗口不会自动将先前的轮次转移到每个 RetrieveAndGenerate 调用中。

A. Query decomposition splits a multi-part question into focused sub-queries, retrieving relevant passages for each part and improving completeness.

B. Temperature affects generation randomness, not how queries are retrieved.

C. A larger model cannot answer using passages that retrieval never returned.

D. Returning fewer results would make incomplete coverage worse, not better.

A. 查询分解将多部分问题拆分为有重点的子查询，检索每个部分的相关段落并提高完整性。

B. 温度影响生成随机性，而不是检索查询的方式。

C. 更大的模型无法使用检索从未返回的段落来回答。

D. 返回较少的结果会使不完全覆盖变得更糟，而不是更好。

A. Knowledge Bases provide fully managed RAG, handling chunking, embeddings, the vector store, and retrieval so the team avoids building the pipeline.

B. A hand-built pipeline maximizes operational overhead, which the team is trying to avoid.

C. Fine-tuning does not keep answers current as documents change and is not managed RAG.

D. Prompt stuffing does not scale and is not a managed retrieval solution.

A. 知识库提供完全托管的 RAG，处理分块、嵌入、向量存储和检索，因此团队可以避免构建管道。

B. 手工构建的管道可以最大限度地提高运营开销，而团队正在努力避免这种情况。

C. 随着文档的更改，微调不会使答案保持最新，并且不受 RAG 管理。

D. 提示填充无法扩展，也不是托管检索解决方案。

A. Semantic chunking groups text by meaning rather than a fixed token count, keeping related content together and improving retrieval relevance.

B. Embedding an entire long document dilutes the vector and harms precise retrieval.

C. Bigger fixed chunks still ignore topic boundaries and can mix unrelated content.

D. Temperature affects generation, not how documents are chunked or retrieved.

A. 语义分块按含义而不是固定的标记计数对文本进行分组，从而将相关内容放在一起并提高检索相关性。

B. 嵌入整个长文档会稀释向量并损害精确检索。

C. 更大的固定块仍然会忽略主题边界，并且可以混合不相关的内容。

D. 温度影响生成，而不影响文档的分块或检索方式。

A. Hierarchical chunking matches on small child chunks for precision and supplies the larger parent chunk as context for generation.

B. One fixed size cannot simultaneously optimize for precise matching and broad context.

C. Treating each document as one chunk loses the precision of small-unit matching.

D. Dimension size does not provide the parent-context behavior hierarchical chunking gives.

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. 一种固定尺寸无法同时针对精确匹配和广泛上下文进行优化。

C. 将每个文档视为一个块会失去小单元匹配的精度。

D. 维度大小不提供分层分块所提供的父上下文行为。

A. Overlap repeats text across chunk boundaries so information spanning a boundary is captured in at least one chunk.

B. Removing overlap makes boundary loss worse, not better.

C. Temperature does not affect chunk boundaries.

D. A larger model cannot recover content that retrieval split across boundaries.

A. Overlap 跨块边界重复文本，因此至少在一个块中捕获跨越边界的信息。

B. 消除重叠会使边界损失变得更糟，而不是更好。

C. 温度不影响块边界。

D. 较大的模型无法恢复检索跨越边界的内容。

A. Query and document embeddings must come from the same model; changing the model requires re-embedding the whole corpus to avoid vector-space mismatch.

B. Vectors from a different model live in a different space and will not match new query embeddings.

C. Mismatched query and document embeddings produce poor or empty retrieval.

D. Temperature is unrelated to embedding compatibility.

A. 查询和文档嵌入必须来自同一模型；更改模型需要重新嵌入整个语料库以避免向量空间不匹配。

B. 来自不同模型的向量位于不同的空间中，并且不会匹配新的查询嵌入。

C. 不匹配的查询和文档嵌入会导致检索效果不佳或为空。

D. 温度与嵌入兼容性无关。

A. Smaller-dimension embeddings reduce index size and query cost, and V2 lets you trade some accuracy for substantial storage savings.

B. Maximum dimensions increase storage and cost, the opposite of the goal.

C. Duplicating vectors increases cost without improving accuracy.

D. Temperature does not affect embedding storage or cost.

A. 较小维度的嵌入可以减少索引大小和查询成本，而 V2 可以让您以一定的准确性换取大量的存储空间节省。

B. 最大尺寸会增加存储和成本，这与目标相反。

C. 复制向量会增加成本，但不会提高准确性。

D. 温度不会影响包埋存储或成本。

A. Returning more candidate chunks raises recall so the correct passage is more likely to be included for generation.

B. Returning fewer results lowers recall and makes misses more likely.

C. Temperature affects generation, not which chunks are retrieved.

D. Arbitrarily shortening the query can remove signal needed for retrieval.

A. 返回更多候选块可以提高召回率，因此正确的段落更有可能被包含在生成中。

B. 返回较少的结果会降低召回率并增加错过的可能性。

C. 温度影响生成，而不影响检索哪些块。

D. 任意缩短查询可以删除检索所需的信号。