AWS Certified Generative AI Developer - Professional (AIP-C01) (AIP-C01) (AIP-C01) — Questions 1–75

Provisioned Throughput reserves dedicated inference capacity for the endpoint, guaranteeing consistent performance regardless of other customers' traffic. This directly satisfies the stem's requirement for a primary advantage over on-demand hosting, where capacity is shared and throughput can be throttled under load.

Bedrock Playground provides a console interface where a single prompt can be submitted to multiple foundation models simultaneously, returning their outputs side-by-side for direct comparison. This satisfies the requirement to evaluate different models against identical input without writing orchestration code.

Streaming returns tokens incrementally as the model generates them, so the first chunk reaches the caller almost immediately rather than after full completion. This directly satisfies the stem's latency constraint: time-to-first-token drops sharply, even though total generation time is unchanged.

Attaching an IAM policy scoped to specific model ARNs satisfies the least-privilege constraint, because Bedrock authorises each InvokeModel call against the resource ARN in the request. Only principals granted bedrock:InvokeModel on those ARNs succeed; other IAM users receive AccessDenied, restricting access to the named foundation models.

SageMaker Data Capture writes JSON Lines records to an Amazon S3 location you specify when enabling the endpoint's data capture configuration, satisfying the requirement to monitor model inputs and outputs. Captured requests and responses land under that prefix, so S3 is the storage destination rather than CloudWatch Logs.

Citation and source attribution let the model tie each generated claim to retrieved passages, which is essential for verifying grounding accuracy. Without attribution, hallucinations cannot be distinguished from grounded statements, so this capability directly satisfies the high-grounding-accuracy requirement in the RAG scenario.

Bedrock Playgrounds provide side-by-side model comparison against the same prompts, and Prompt Management stores and versions those prompts, removing custom scripting. This directly satisfies the stem's requirement to compare Claude 3 Sonnet with Llama 3 on a defined prompt set.

Bedrock Model Invocation Logging captures each request and response, including the model ID and version invoked, delivering records to CloudWatch Logs or S3. This provides the per-request audit trail needed to identify the exact foundation model version used for every inference.

Filtering on document metadata during retrieval enforces per-user authorisation at query time, satisfying the requirement that users only see documents they are permitted to access. Metadata tags such as department or group are matched against the caller's identity, so unauthorised chunks never enter the context passed to the model.

Provisioned throughput is economically justified when sustained token consumption is high enough that the committed rate beats on-demand pricing and the guaranteed capacity prevents throttling. With near-zero overnight traffic, a 24/7 commitment pays for idle hours. Running on-demand during steady business hours and reserving only when utilization analysis shows a durable break-even point avoids paying for capacity nobody uses.

Application Auto Scaling scheduled actions adjust the endpoint's desired instance count against a cron or rate schedule, scaling in during off-peak hours and out afterwards. This directly satisfies the off-peak cost-reduction constraint without manual intervention or redeploying the endpoint.

Trusted Advisor surfaces cost-optimisation findings across the account, flagging idle or underutilised SageMaker endpoints, notebooks and training instances. This directly satisfies the stem's requirement to identify savings opportunities in the existing environment without deploying new tooling.

Knowledge Bases assume an IAM service role to read source documents from S3 and invoke embedding models, so permissions are scoped to that role rather than embedded credentials. This satisfies least-privilege access to the underlying data stores.

Guardrails support custom word filters that can be configured to block specific terms or phrases. By adding the subtle hate speech patterns to a custom filter, the developer can increase detection without broadly raising sensitivity, thus minimizing false positives on benign content. This is a targeted approach.

InvocationsPerInstance measures the request load each endpoint instance actually handles, so scaling policies trigger on real per-instance demand rather than aggregate traffic. This directly satisfies the stem's requirement for a meaningful auto-scaling metric, since it reveals when individual instances are saturated and additional capacity is needed.

SageMaker JumpStart vision models expect image bytes serialised as a Base64 encoded string within the JSON input field, not raw binary or a URL. Supplying the image in that encoding resolves the schema mismatch at InvokeEndpoint.

Validating the reasoning path verifies the agent's chain-of-thought and orchestration logic, not just the final response. This satisfies the test plan requirement to assess how the agent reaches conclusions, exposing flawed planning or prompt interpretation before production deployment.

ThrottlingException occurs when request rate or token throughput exceeds the account's quota. Exponential backoff with jitter lets retries succeed when the limit is transient, while a quota increase raises the ceiling for sustained peak load. Together they address both the immediate retry behavior and the underlying capacity limit without altering the model or application logic.

Guardrails offers sensitive information filters that detect and mask or block PII and PHI patterns in prompts and responses, plus prompt attack filters that detect jailbreak and prompt-injection attempts aimed at extracting system instructions. Enabling it addresses both concerns at the model boundary without application changes.

Metadata filtering applies attribute-based predicates during retrieval, so each user's query only returns chunks whose metadata matches their identity claims. This enforces per-user access control at query time without duplicating knowledge bases or pre-filtering the entire corpus.

Configuring the vector database with KMS encryption applies customer-managed keys to the underlying storage volumes, ensuring chunks are encrypted at rest. This satisfies the encryption-at-rest constraint directly, rather than relying on transport TLS or application-layer tokenisation.

Foundation models are large pre-trained neural networks that underpin generative tasks in Amazon Bedrock, producing text completions, summaries and analysis from prompts. This satisfies the stem's requirement to identify their purpose: enabling generative output rather than classification or embedding-only workloads.

Fine-tuning adjusts model weights on curated examples, which reliably instils a consistent house style, vocabulary and formatting that prompting or RAG cannot guarantee. RAG supplies external knowledge at inference time but leaves tone unchanged, so it cannot satisfy this stylistic requirement.

A Knowledge Base performs retrieval-augmented generation: documents are chunked and embedded, relevant chunks are retrieved per query, and only those passages are passed to the model. This satisfies the constraint that source documents exceed the prompt size limit.

AWS Budgets lets you define a monthly cost budget and configure alerts when actual or forecasted spend crosses your threshold. This directly satisfies the requirement to be notified before the SageMaker inference endpoint exceeds its specified monthly budget.

Inference latency measures the time the RAG system takes to return a response, capturing retrieval plus generation overhead. As concurrent usage grows, this metric reveals degradation from model contention or vector store load, satisfying the need to track performance under increasing demand.

Claude 3 Sonnet accepts both text and image inputs, satisfying the multimodal requirement. Within Amazon Bedrock, it handles vision tasks such as chart and document analysis alongside text prompts, unlike text-only models such as Titan Text. This makes it the appropriate choice when a single model must process mixed image and text content.

The embedding model encodes text chunks and queries into dense numerical vectors, placing semantically similar content close together in vector space. This enables the vector store's similarity search to retrieve relevant context for the generator.

Embedding models convert text into vector representations and are offered as fully managed, generally available foundation models within Amazon Bedrock. They support RAG and semantic search workloads, satisfying the question's requirement for a model type generally available in the service.

Amazon Titan Text Embeddings is natively integrated with Amazon Bedrock Knowledge Bases, so vectors are generated and indexed without custom embedding pipelines or dimension mismatches. That native fit delivers the seamless integration and scalability the scenario requires.

Switching to a compute-optimised instance type directly addresses the CPU-bound constraint by providing a higher ratio of vCPU to memory, such as the C-family. This supplies the processing capacity the model needs during inference, relieving the bottleneck that general-purpose instances impose.

Limiting IAM permissions for the fine-tuning job enforces least privilege, so the job can read only its training data and write only its output model. This constrains blast radius if the job is compromised, satisfying the security requirement for scoped access during customisation.

Amazon Bedrock Agents parse the Lambda response according to the OpenAPI schema attached to the action group. If the function returns an unexpected top-level key, the agent treats the invocation as malformed and stops the orchestration. Aligning the Lambda output with the schema's response body structure restores successful parsing and lets the agent continue the multi-step task.

Amazon OpenSearch Serverless vector engine provides the low-latency similarity search the knowledge base demands, and its native LangChain integration satisfies the stated requirement. It scales automatically for large embedding volumes without capacity planning, unlike alternatives lacking managed serverless scaling or first-class LangChain support.

Hierarchical chunking, such as using a parent-child relationship, preserves the document structure and allows retrieval of smaller, more precise child chunks while providing the parent chunk for context. For technical manuals with code snippets and paragraphs, this approach improves retrieval accuracy by matching queries to relevant sections without losing context.

S3 Lifecycle policies automatically transition or expire pipeline artefacts after defined periods, moving older outputs to cheaper storage classes or deleting them. This directly addresses the stem's storage-cost optimisation goal without altering pipeline logic or retraining workflows.

Amazon Bedrock Knowledge Bases natively ingest unstructured documents from Amazon S3, supporting PDF, Markdown, HTML, and plain text through the built-in parsing pipeline. This satisfies the stem's ingestion requirement directly, since no custom chunking or transformation layer is needed before vectorisation into the underlying vector store.

A Guardrail with a Denied Topics policy blocks the medical-advice subject outright, matching the requirement that the application must not provide medical advice. The policy intercepts prompts and responses on that topic before they reach the user, enforcing the constraint at the model boundary.

Invocations counts requests hitting the endpoint, exposing traffic volume and unexpected drops that signal throttling, client errors or misrouting. CloudWatch publishes it automatically for SageMaker endpoints, making it a primary operational-health signal for detecting demand shifts and failures.

Few-shot prompting embeds a small set of exemplar input-output pairs directly in the prompt, letting the model infer the task pattern through in-context learning without weight updates. This satisfies the stem's requirement for guiding generation via demonstrated examples rather than instructions alone.

Guardrails provides a denied topics policy that blocks investment-advice content and sensitive information filters for PII, while model invocation logging captures the full prompt, response, and model ID for traceability. Both are configured at the account or application level and apply without rewriting each application's code.

The Base Prompt templates how retrieved chunks are injected and instructs the model to answer strictly from that context, satisfying the Knowledge Base's grounding requirement. It governs response formatting and citation behaviour, not retrieval itself — chunking and vector search are configured separately. This keeps generated answers anchored to your indexed data rather than the model's pretrained knowledge.

Guardrails for Amazon Bedrock allows you to define policies that filter and block harmful or inappropriate content based on categories such as hate, violence, and sexual content. It can also be configured to log blocked content for auditing. This directly addresses the requirement to prevent harmful content and maintain an audit trail.

The Amazon Bedrock documentation region table lists exactly which foundation models are available in each AWS region, directly satisfying the data residency constraint. Model availability varies by region, so this table is the authoritative source for confirming that a chosen model can be hosted within the required jurisdiction.

A Multi-Model Endpoint hosts several models behind shared inference containers, loading them on demand from Amazon S3. Deploying a JumpStart model this way avoids provisioning a dedicated endpoint per model, satisfying the resource utilisation constraint by sharing compute across hosted models.

Bedrock Guardrails applies configurable PII detection and masking policies directly to prompts and completions before the foundation model processes them, satisfying the requirement that sensitive data be filtered prior to inference rather than relying on post-processing or application-side scrubbing.

Content filtering evaluates prompts and responses against configurable harmful-category thresholds, such as hate, violence and sexual content, and blocks them when exceeded. This satisfies the question's requirement for a guardrail capability that screens model input and output for harmful material.

A temperature of 0.0 makes decoding deterministic and strongly favours the highest-probability tokens, while system instructions constrain the model to answer only from retrieved context. Together these reduce unsupported generation, keeping output grounded in the provided documents.

Multi-region deployment with cross-region replication directly satisfies the resilience requirement by removing single-region failure as a blast radius. If one region's vector database or inference endpoints become unavailable, traffic fails over to a replica holding synchronised embeddings, keeping retrieval and generation operational. Microsoft Entra ID governs access across both regions without reconfiguration.

Bedrock throttling is transient and retrying with exponential backoff and jitter is the AWS-recommended approach. The SDK's built-in retry mechanism can be customized, and adding jitter prevents synchronized retries from multiple Lambda instances. This directly addresses the throttling without changing quotas.

AWS Service Terms govern the platform-level agreement, while each model provider publishes its own data processing terms covering how inputs and outputs are handled. Reviewing both sources is necessary because provider-specific policies supplement, and may differ from, the baseline AWS terms.

TLS 1.2+ encrypts data in transit across every ingestion hop — source systems, S3 uploads, and vector store writes — satisfying the pipeline's transport-encryption requirement. This prevents interception or tampering of documents and embeddings moving between services, complementing at-rest encryption rather than replacing it.

An explicit IAM policy denying the DeleteKnowledgeBase action overrides any allow granted by attached identity policies, directly preventing accidental deletion. This satisfies the stem's constraint of blocking deletion while leaving other Knowledge Base operations unaffected.

RAGAS quantifies retrieval-augmented generation quality by scoring faithfulness (whether claims are entailed by retrieved context) and answer relevance, directly measuring ground-truth accuracy against the internal knowledge base. This satisfies the stem's requirement to quantify hallucination reduction in the RAG pipeline.

Amazon CloudWatch captures Bedrock invocation metrics and logs, including model usage, token counts and latency, which can be aggregated into a usage report. It satisfies the auditor's requirement by providing operational telemetry across the account.

Serverless Inference scales to zero when no requests arrive, so you pay only for compute during actual invocations. This directly addresses the unpredictable traffic and significant idle periods, unlike real-time endpoints that bill for continuously provisioned instances regardless of load.

Dynamic batching groups concurrent inference requests into a single forward pass, raising GPU utilisation and cutting per-request latency. It directly addresses the low memory utilisation and high latency described, since the GPU is otherwise underfed by one-at-a-time requests. This suits real-time endpoints with concurrent traffic.

An action group defines the API operations the agent may invoke, including the OpenAPI schema describing parameters and the Lambda function that fulfils each call. It is the required binding that lets Bedrock Agents interact with the external API.

Setting temperature to 0 makes the model's output more deterministic by reducing randomness in token selection. While not perfectly deterministic due to floating-point operations, it is the primary parameter to control variability. This is the correct adjustment for consistent responses across identical prompts.

Naming the frameworks or libraries constrains the assistant to idiomatic APIs and imports rather than generic standard-library substitutes, so generated code integrates with the project's existing dependencies — one of the three prompt elements the stem requires for best results.

A restrictive system prompt constrains generation to retrieved context, instructing the model to refuse or state uncertainty when evidence is absent. This directly addresses hallucination by preventing the model from drawing on unsupported parametric knowledge outside the supplied documents.

Specifying subnets and security groups during model creation places the inference container's network interfaces inside the VPC, giving it private connectivity and the correct security group rules. Without this configuration, the endpoint cannot reach private VPC resources.

Managed Spot Training taps spare AWS capacity at up to 90% discount, satisfying the cost-reduction constraint. SageMaker manages interruptions by checkpointing to Amazon S3 and resuming, so long jobs tolerate reclaimed instances without restarting from scratch.

A secondary validation model inspects prompts for malicious intent before inference, catching injection attempts that static keyword filters miss because attackers rephrase payloads. This layered detection satisfies the prompt-injection protection requirement for the SageMaker JumpStart endpoint more effectively than input sanitisation alone.

CloudWatch metrics expose Bedrock's invocation-level telemetry, including 5xx error rates and throttling counts, which surface anomalous request patterns indicative of abuse. Satisfying the monitoring constraint, these native service metrics feed alarms and dashboards without custom instrumentation, enabling prompt detection of volumetric or error-driven misuse.

Interface VPC endpoints (PrivateLink) keep Bedrock traffic on the AWS private network, avoiding public internet exposure entirely. This satisfies the secure perimeter requirement by giving private connectivity from the VPC, with endpoint policies and security groups governing access.

toolConfig is the correct parameter to define tools.

Auto Scaling policies driven by custom metrics such as invocation latency or queue depth scale instance counts to match real traffic, avoiding over-provisioned idle capacity while preserving throughput during peaks, which directly reduces endpoint cost.

Selecting an embedding model is mandatory because Amazon Bedrock Knowledge Bases vectorise your source chunks with it before indexing into the vector store. The embedding model defines the vector space and dimensionality, so it must be chosen at creation and cannot be swapped later without re-ingesting the entire data source.

Regular model evaluation for bias detects disparate performance across demographic groups before and after deployment, enabling remediation. This directly satisfies the stem's responsible AI requirement by providing measurable, recurring evidence that outputs remain fair throughout the production workload's lifecycle.

AWS PrivateLink (VPC Endpoints) keeps traffic to Amazon Bedrock entirely within the AWS network, satisfying the private connectivity constraint. Interface endpoints use ENIs in your subnets, so requests never traverse the public internet, and TLS encrypts data in transit between your VPC and the service.

Provisioned Concurrency pre-warms resources to avoid cold starts.

AWS states that customer inputs and outputs for third-party models on Bedrock are not used to train the underlying foundation models, and are not shared with model providers. Recognising this satisfies the stem's requirement without needing opt-out configuration, since the protection applies by default.

The system parameter supplies a system-level instruction that establishes the model's role or persona before user turns are processed. It satisfies the requirement to define persona at the API level, distinct from user messages and other inference configuration parameters.

Model Invocation Logs capture the full request and response payloads—prompts, completions, and metadata—that Bedrock passes to and from a model, then deliver them to CloudWatch Logs or Amazon S3. This satisfies the auditing requirement by preserving an inspectable record of every invocation for compliance and troubleshooting.