After a prompt update, a Bedrock application's average input length triples while request counts remain unchanged. Traffic bursts now cause throttling, and synchronized retries produce additional spikes.
Measured average token demand still fits the model's applicable quota. Requests can tolerate brief queuing within their existing deadlines. The application must stabilize the regional deployment while a quota review is pending.
Which TWO changes should it prioritize?
Increase worker concurrency and retain the existing request-count admission limit.
Use a fixed retry interval across workers and cap each worker's retry count.
Pace admission using estimated token demand and the model's applicable quota accounting.
Apply bounded exponential backoff with jitter, stopping retries when the request deadline is exhausted.
C
Token-aware pacing addresses the changed request cost, and jittered backoff reduces synchronized retry bursts while respecting deadlines. Raising concurrency under a request-count limit increases pressure, and a fixed retry interval retains synchronization. These controls smooth bursts; they do not create additional quota. The question establishes that average demand fits available capacity.
Testing, Validation, and Troubleshooting
25
Select Two
D
A candidate prompt improves the overall score on the dataset used during prompt development. Reviewers suspect it performs worse on uncommon but business-critical requests.
The team has enough labeled examples to evaluate those categories separately and can route a small portion of production traffic to a candidate version.
Which release process provides the strongest evidence?
Evaluate held-out category slices, apply critical-category acceptance criteria, and canary with predefined rollback thresholds.
Evaluate held-out data, apply an aggregate acceptance threshold, and canary while collecting overall user ratings.
Re-evaluate the development data by category, approve the critical-category results, and canary with rollback thresholds.
Evaluate held-out category slices, approve the overall score improvement, and release globally with incident monitoring.
A
Held-out data tests generalization, and critical-category criteria prevent aggregate improvements from hiding unacceptable regressions. A canary then limits exposure and supplies predefined rollback decisions. An aggregate threshold ignores the category risk, reusing development data does not test generalization, and a global release lacks controlled rollout despite the identified category risk.
Testing, Validation, and Troubleshooting
24
Select One
A RAG release changes both retrieval configuration and the generation prompt. End-to-end answer quality declines.
The team has a held-out dataset containing questions, approved evidence passages, and answer-quality criteria. It needs to determine whether to revert retrieval, the prompt, or both.
Which TWO experiments best isolate the components?
Compare both releases end to end and group their final-answer scores by question category.
Compare each retrieval configuration's returned passages against the approved evidence, without generation.
Ask an evaluator model to assign each incorrect answer to retrieval or generation using the answer text alone.
Run both generation prompts against identical approved passages and compare their answers under the same criteria.
B
Comparing retrieved passages with the approved evidence tests whether retrieval finds the required evidence, and running both prompts against identical approved passages holds evidence constant while comparing generation. End-to-end comparisons measure combined effects, and an evaluator model working from the answer text alone lacks the retrieved evidence needed for a reliable component-level diagnosis.
Testing, Validation, and Troubleshooting
23
Select Two
D
An application currently uses a large model for every request. Evaluation shows that a smaller model meets quality requirements for routine requests but not for complex requests.
A separate classifier identifies routine cases with acceptable error rates. It also produces an uncertain category. A small-model attempt followed by a large-model fallback would exceed the response deadline.
Which design best reduces model expenditure while preserving the measured quality and latency requirements?
Send requests to the smaller model and invoke the larger model when the generated answer reports uncertainty.
Run both models concurrently for every request and use the classifier to select which response to return.
Send routine and uncertain requests to the smaller model and reserve the larger model for confirmed complex cases.
Send classified routine requests to the smaller model and route complex or uncertain requests directly to the larger model.
D
The measured classifier supports using the smaller model for routine requests. Routing uncertain cases directly to the larger model avoids an unvalidated quality assumption and sequential fallback latency. A small-model attempt with fallback violates the latency constraint, running both models retains the large-model expense for every request, and sending uncertain cases to the smaller model uses the less capable path.
Operational Efficiency and Optimization for GenAI Applications
22
Select One
A company summarizes documents with a Bedrock model. Most documents arrive in a weekend backlog and are needed the following day. A small number are submitted by support staff and must be returned within seconds.
Tests show that batch inference meets the backlog's deadline at lower cost than individual calls. On-demand invocation meets the interactive latency requirement.
Which processing design best meets both requirements?
Place both request types in the batch input and prioritize support documents at the beginning of each input file.
Submit the backlog through batch inference and route interactive requests through on-demand invocation.
Process both request types through on-demand workers and lower the backlog's worker priority during support traffic.
Run support requests against both paths and return the first completed result while retaining batch processing for the backlog.
B
The measured results support different invocation paths for different deadlines. Putting support documents first in a batch input does not establish interactive completion guarantees. Processing everything on demand gives up the demonstrated backlog savings, and duplicating interactive requests across both paths adds cost without a demonstrated requirement or benefit. The cost and latency conclusions come from the scenario's tests, not an assumption that every model or workload behaves identically.
Operational Efficiency and Optimization for GenAI Applications
21
Select One
A Bedrock application repeatedly sends the same long policy document with different user questions. The selected model supports explicit prompt caching, and the shared content satisfies the model's cache-size requirements.
Requests normally arrive within the cache lifetime. Each question still requires a newly generated answer.
Which arrangement is most likely to maximize reuse of the shared prompt content?
Place the user identifier and question first, followed by the policy document and a cache checkpoint.
Place the policy first, append the changing question, and put a cache checkpoint after the complete prompt.
Place the policy first, put a cache checkpoint after it, and append the changing question.
Place a request timestamp before the policy, put a cache checkpoint after the policy, and append the question.
C
The checkpoint covers the stable prefix while the changing question follows it. The other arrangements place changing content before the relevant checkpoint or inside the cached prefix, preventing consistent reuse of that prefix. Prompt caching reuses prompt processing; the application still generates a new answer.
Operational Efficiency and Optimization for GenAI Applications
20
Select One
A company invokes supported Bedrock runtime operations from several AWS accounts and Regions. Incident investigators need full prompt and response payloads in a central security account.
The team wants to use native invocation logging and minimize application changes. Source-account storage is permitted if access and retention are controlled.
Which design meets these requirements?
Configure invocation logging to local S3 destinations and replicate the resulting objects to the central security account.
Configure invocation logging in every account to write directly to one S3 bucket in the central security account.
Configure invocation logging in every Region to write directly to one regional S3 bucket in each source account.
Configure an organization CloudTrail trail and use its centralized event records as the prompt-and-response archive.
A
Native invocation-logging destinations must be in the same account and Region as the logging configuration. Local destinations followed by S3 replication accommodate central collection. Writing directly to a central-account bucket violates the account constraint, and a single regional bucket in each source account violates the Region constraint for other Regions. CloudTrail does not replace full invocation-payload logging.
AI Safety, Security, and Governance
19
Select One
An EC2 application invokes Bedrock through the standard regional bedrock-runtime hostname. The VPC has no internet egress, and the application must retain its current SDK endpoint configuration.
The instance role currently allows invocation of all models. The security team requires private connectivity and access limited to an approved model list.
Which configuration meets both requirements?
Create a runtime interface endpoint with private DNS and restrict access through its security group to the application instances.
Create a control-plane interface endpoint with private DNS and narrow the instance role to the approved runtime model resources.
Create a runtime interface endpoint without private DNS and narrow the instance role to the approved model resources.
Create a runtime interface endpoint with private DNS and constrain model access through the instance role and endpoint policy.
D
The runtime endpoint supplies the inference path, while private DNS preserves the standard hostname. IAM and endpoint policies can constrain model access. A security group alone does not restrict models. A control-plane endpoint is the wrong service endpoint. Without private DNS, the unchanged standard hostname does not resolve to the private endpoint.
AI Safety, Security, and Governance
18
Select One
A Bedrock application streams responses to users. A pilot using asynchronous guardrail processing sometimes displays text before a guardrail intervention arrives.
The application must continue delivering responses incrementally, but generated text must undergo the configured guardrail checks before display. The selected model and streaming API support this behavior.
Which change best meets the requirement?
Retain asynchronous processing and add a short fixed client-side delay before displaying each chunk.
Use synchronous guardrail processing and forward chunks after the service completes their checks.
Retain asynchronous processing and remove earlier displayed chunks when an intervention is received.
Buffer the full response, evaluate it with ApplyGuardrail, and send the approved text to the client.
B
Synchronous processing buffers and checks response chunks before forwarding them. A fixed delay does not establish that checks have finished, and removing chunks after an intervention permits exposure before it arrives. Buffering the full response checks before display but loses incremental delivery. This establishes evaluation order, not perfect detection of every policy violation.
AI Safety, Security, and Governance
17
Select One
A multi-tenant assistant uses a shared vector store. Documents receive a trusted tenant_id during ingestion. The assistant can also call backend tools to read and update tenant-owned records.
Users control their questions, and retrieved documents can contain malicious instructions. The application service is the only component allowed to access the vector store.
Which TWO controls establish tenant boundaries independently of model behavior?
Build a mandatory retrieval filter from authenticated tenant context and prevent requests from overriding it.
Let the model select the tenant filter, then validate its value against the list of tenants configured in the application.
Authorize each tool operation against the authenticated caller and the ownership of the requested resource.
Grant tools a service identity and authorize operations using the tenant identifier supplied in the generated tool arguments.
A
Trusted application identity must determine retrieval scope and tool authorization. Letting the model choose the tenant filter only verifies that a tenant exists, not that the caller belongs to it, and trusting tenant identifiers in model-generated tool arguments substitutes model output for the authenticated caller.
AI Safety, Security, and Governance
16
Select Two
C
An application uses both Bedrock-hosted models and an external model provider. An approved Bedrock guardrail configuration defines content restrictions for user inputs and generated outputs.
The external provider must remain in use. Both providers must apply the approved guardrail version before prohibited content is forwarded.
Which integration best meets these requirements?
Attach the guardrail to Bedrock calls and reproduce its settings in the external provider's moderation configuration.
Evaluate external-model outputs with the guardrail and use the provider's input moderation for external requests.
Apply the approved guardrail version to external inputs and outputs through ApplyGuardrail, handling interventions in the application.
Generate external responses first and pass them through a Bedrock model configured to rewrite content under the guardrail.
C
ApplyGuardrail can assess text independently of model invocation. Using the approved version for both external inputs and outputs preserves the configured policy boundary. Reproducing settings or relying on the provider's input moderation substitutes different moderation behavior. Rewriting after external processing acts too late and introduces rewriting rather than enforcing the required input-and-output flow.
AI Safety, Security, and Governance
15
Select One
A Step Functions Standard Workflow pauses for a human decision before publishing a generated assessment. Review can take 48 hours.
Reviewers use an existing application. They must authenticate, and each approval must apply to the specific assessment version being reviewed. Workflow task tokens must not be exposed in browser URLs.
Which design best meets these requirements?
Email the execution ARN and let the application start a new publication execution after authenticating the reviewer.
Store the assessment status and use a scheduled task to restart the workflow when the application marks it approved.
Send the task token in an encrypted approval URL and let the browser submit it after the reviewer signs in.
Store the task token server-side against the review record and submit the callback after authorization and version checks.
D
A callback task can pause the Standard execution until the application submits its result. Server-side association lets the application check reviewer authorization and assessment version without exposing the token in a browser URL. Starting a new execution or restarting the workflow introduces separate execution coordination, and putting the token in a URL violates the token-handling requirement.
Implementation and Integration
14
Select One
An AgentCore Gateway exposes a supported external REST API as an MCP tool. Agents authenticate to the gateway using JWTs from the company's identity provider.
The external API uses a different OAuth provider and expects an application credential. Its required OAuth flow is supported by the gateway. Valid agent requests reach the gateway, but target calls fail with authorization errors.
Which change best addresses the failure?
Add the external API's audience to the inbound JWT configuration and reuse the incoming agent token for target access.
Configure the target's outbound OAuth credentials while retaining the gateway's existing inbound JWT authorization.
Add the external API's OAuth scopes to the inbound authorizer and keep the target's credential configuration unchanged.
Grant the gateway execution role additional AWS invocation permissions and retain the current external API authentication.
B
Inbound authorization governs agent access to the gateway, while outbound credentials authenticate the gateway to its target. Changing inbound audiences or scopes does not supply the target's required OAuth credential, and AWS IAM permissions do not replace the external API's OAuth authentication.
Implementation and Integration
13
Select One
An application hosts an agent in AgentCore Runtime. A user can open several independent investigations, each involving multiple requests and temporary working files. Requests within one investigation should reuse its runtime context, while different investigations must remain isolated.
Durable results are stored outside the runtime.
Which session-management design best meets these requirements?
Assign a session identifier to each investigation, verify its ownership, and reuse it for that investigation's requests.
Assign a session identifier to each user and separate concurrent investigations through prompt labels and working-directory names.
Generate a session identifier for each model turn and pass the previous turn's session identifier as conversation metadata.
Assign a session identifier to each tenant and serialize requests while clearing temporary files between investigations.
A
A runtime session provides an isolated execution context. Reusing its identifier supports follow-up interactions within that context; ownership checks remain necessary. Per-user and per-tenant sessions combine independent investigations. A session per turn creates separate environments rather than retaining runtime context. Session state remains ephemeral; external storage is needed for durable results.
Implementation and Integration
12
Select One
A document-analysis service performs dependent model calls, validation, and occasional manual review. A job can remain active for several hours.
The submission endpoint must return a job identifier promptly. Processing must survive client disconnection, and the team wants execution history and recovery without building a separate workflow engine.
Which architecture best meets these requirements?
Start an Express Workflow asynchronously and use a DynamoDB item to retain status throughout the job.
Start a Standard Workflow synchronously through the submission endpoint and return its result as the job response.
Start a Standard Workflow asynchronously and expose job status through a separate authenticated endpoint.
Start short Express executions for each phase and implement phase transitions and review coordination in a custom service.
C
An asynchronously started Standard execution supports the long-running workflow while the submission endpoint returns promptly. An Express Workflow does not overcome its duration limits. Standard workflows do not support the synchronous Express invocation pattern implied by the synchronous option. Chaining short Express executions can be engineered but adds custom orchestration and recovery.
Implementation and Integration
11
Select One
A Bedrock agent collects refund details. Refund approval and execution must remain in an existing business application. The application records approval for a specific customer, order, and amount.
A timeout can occur after the payment system processes a refund but before the application receives the response.
Which TWO controls best preserve the approval boundary and prevent duplicate refunds?
Enable conversational confirmation and map an affirmative response to the application's approved-refund status.
Return control to the application and execute only the transaction parameters covered by its recorded approval.
Generate a new refund identifier after a timeout and reconcile duplicate payments during the settlement process.
Use a backend idempotency key tied to the approved logical refund and reuse it when retrying that transaction.
B
Return control allows application-owned approval and execution, and binding execution to the approved parameters prevents changes after approval. A stable logical-refund key lets the backend deduplicate retries. Conversational confirmation is not the required recorded approval, and generating new retry identifiers defeats deduplication.
Implementation and Integration
10
Select Two
D
An application uses the Bedrock Converse API with client-side tool use. The model requests an order-status lookup through a toolUse block.
The application validates the arguments, confirms the caller can access the order, and executes the lookup. It must now allow the model to complete its answer using the result.
What should the application send next?
A new conversation containing the original question and the lookup result as a replacement system instruction.
The existing conversation with the lookup result appended as ordinary user text identifying the tool by name.
The existing conversation with an assistant text message containing the lookup result and the order identifier.
The conversation including the assistant's tool request, followed by a user message containing a matching toolResult.
D
Converse client-side tool use requires the application to return a structured tool result associated with the model's tool-use request. Preserving the request and its matching identifier lets the model continue the exchange. The other options replace or bypass that protocol with ordinary conversation content.
Implementation and Integration
9
Select One
An assistant uses an S3-backed Bedrock knowledge base and caches generated answers. Editors update and withdraw source documents in controlled refresh windows.
The application pauses answer generation during a refresh. Before reopening, it must ensure that responses use the refreshed corpus and cannot reuse answers from the previous corpus version. Full re-embedding of unchanged documents should be avoided.
Which process best meets these requirements?
Synchronize source changes, verify the updated corpus is query-visible, and advance the answer-cache version before reopening.
Synchronize source changes, reopen when the ingestion job reports completion, and expire cached answers after their normal TTL.
Advance the answer-cache version, reopen after S3 updates complete, and synchronize the knowledge base in the background.
Verify source-file checksums, synchronize changed files, and retain cached answers whose cited document identifiers still exist.
A
The process must account for both retrieval visibility and cached answers. AWS notes that query visibility can lag ingestion completion for some vector stores. Reopening on job completion overlooks that lag and retains old answers. Reopening before synchronization exposes the old corpus. Keeping answers whose cited document identifiers still exist does not detect changed content under an unchanged identifier. Cache versioning and the paused refresh boundary are application controls in this scenario.
Foundation Model Integration, Data Management, and Compliance
8
Select One
A Bedrock application needs additional inference capacity. The organization permits processing in an approved set of European Regions.
A geographic cross-Region inference profile contains only approved destinations. The application role currently permits the foundation model in the source Region, but organization policies deny some of the profile's destination Regions.
Which TWO changes are required to use this profile within the approved boundary?
Authorize the profile and the required foundation-model resources in its source and destination Regions.
Keep source-Region model permissions and use the profile identifier to delegate destination authorization to Bedrock.
Amend the Region restrictions to permit the required inference routing while retaining the organization's boundary.
Grant access to the profile resource and use the source endpoint's Region as the processing-location restriction.
A
Authorization must cover the inference profile and the required model resources in its source and destination Regions, and the organization's Region restrictions must also permit the profile's routing while keeping the approved boundary. Selecting a profile does not delegate authorization checks to Bedrock, and the source endpoint's Region alone does not constrain where processing occurs.
Foundation Model Integration, Data Management, and Compliance
7
Select Two
C
Three applications share a prompt in Bedrock Prompt Management. Their deployment process currently retrieves the working draft. A recent edit improved one application but caused regressions in another.
Prompt authors must continue experimenting, while application owners need independent releases and rollback.
Which process best meets these requirements?
Give each application its own working draft and copy approved edits into those drafts before deployment.
Publish tested prompt versions and configure each application to reference its approved version.
Select the latest published version at startup and record its identifier with each application release.
Resolve the shared draft during deployment and retain the prompt identifier in the release manifest.
B
Published prompt versions are snapshots that applications can select independently. Per-application drafts retain mutable production references. Selecting the latest version at startup lets a restart adopt a version that an application owner has not approved. Recording the prompt identifier does not pin the mutable draft's contents.
Foundation Model Integration, Data Management, and Compliance
6
Select One
A policy assistant frequently retrieves the clause relevant to a question but misses exceptions contained elsewhere in the same section.
An evaluation shows that 250-token chunks provide better matching than 1,500-token chunks. However, answers are more accurate when the generator receives the full relevant section. The application uses Bedrock Knowledge Bases, and the team wants to avoid maintaining a custom two-stage retrieval service.
Which change is the best starting point?
Use 1,500-token fixed-size chunks and increase retrieval depth to compensate for weaker matching.
Use overlapping 250-token chunks and rely on a larger retrieval limit to collect nearby exceptions.
Use hierarchical chunking with smaller child chunks and parent chunks sized to retain the relevant section context.
Use semantic chunking and return the highest-ranked individual chunks within the existing context budget.
C
Hierarchical chunking retrieves smaller child chunks and substitutes broader parent context. It directly addresses the measured precision and context trade-off. Larger fixed chunks sacrifice the better small-chunk matching. Overlapping small chunks may retrieve nearby exceptions but do not ensure them. Semantic chunking changes boundaries without establishing the required parent context.
Foundation Model Integration, Data Management, and Compliance
5
Select One
An S3 collection contains text-only policy PDFs and illustrated maintenance manuals. The default Bedrock Knowledge Bases parser produces satisfactory results for the policies. Evaluation shows that the manuals need extraction of figures and tables.
The team has validated Bedrock Data Automation parsing for the manuals. It wants to avoid paying for advanced parsing of the text-only policies while keeping both collections searchable through the application.
Which ingestion design best meets these requirements?
Separate the collections into data sources and configure the parser independently for each source.
Use one data source with Data Automation parsing and identify text-only files through document metadata.
Use one data source with foundation-model parsing and instruct the parser to skip visual analysis for policies.
Use one data source with Data Automation parsing and configure smaller chunks for text-only policies.
A
Parser configuration applies at the data-source level. AWS documents that an advanced parser processes all PDFs in that source, including text-only PDFs. Separating sources allows different parsing choices. Metadata, extraction instructions, and chunk sizes do not make advanced parsing free for selected PDFs.
Foundation Model Integration, Data Management, and Compliance
4
Select One
A custom RAG application uses embedding model A for documents and queries. The team selects model B after an offline evaluation. Both models produce vectors of the same dimensionality, but their embedding spaces were learned independently.
The migration must preserve search availability and allow a quick return to the previous configuration.
Which deployment plan best meets these requirements?
Change query embedding to B first, then replace document embeddings in the existing index during normal ingestion.
Update document embeddings in place, retaining each original vector in an audit store for restoration if required.
Add B-generated documents to the existing vector field and use a model-version filter while backfilling the collection.
Populate a separate B-based index, validate queries using B, and switch the query-model and index configuration together.
D
Equal dimensionality does not make independently learned embedding spaces interchangeable. A separate index preserves the previous configuration while B-based retrieval is validated. Changing query embeddings first mixes query and document spaces, in-place updates require restoring overwritten vectors, and adding B vectors to the existing field leaves the B-filtered corpus incomplete during backfill.
Foundation Model Integration, Data Management, and Compliance
3
Select One
A support assistant uses Amazon Bedrock Knowledge Bases with an OpenSearch Serverless vector store configured to support hybrid search.
Evaluation identifies two failure patterns. Exact product-code queries sometimes fail to retrieve the matching document. For descriptive queries, the required evidence often appears within the first 40 candidates but falls outside the five passages supplied to the generator.
The team can increase retrieval-stage processing, but generation must remain limited to five passages.
Which TWO changes most directly address these findings?
Increase chunk size so product identifiers and broader descriptive content are embedded together.
Apply an exact product-code metadata filter to every query before performing semantic retrieval.
Combine lexical and semantic retrieval by configuring hybrid search for the affected queries.
Rerank the candidate passages against the original query and pass the highest-ranked five.
C
Hybrid search addresses lexical identifiers such as product codes alongside semantic similarity. Reranking improves selection from the candidate pool without expanding the five-passage generation context. Larger chunks may reduce matching precision, and a metadata filter on product code cannot handle descriptive queries that contain no explicit code.
Foundation Model Integration, Data Management, and Compliance
2
Select Two
D
A company's assistant answers questions about internal procedures. Its current foundation model meets language and formatting requirements, but evaluation shows that answers sometimes rely on outdated procedures.
The authoritative repository changes daily. Reviewers must be able to inspect the passages used for each answer. The company can operate an ingestion pipeline but has limited capacity to maintain model-customization jobs.
Which approach best meets these requirements?
Fine-tune weekly on approved procedures and retrieve document titles separately to populate the answer's references.
Retrieve passages from an incrementally refreshed index and generate answers using those passages and their source identifiers.
Generate daily departmental summaries and select the relevant summary as the model's system context for each request.
Cache approved answers by semantic similarity and refresh each cached answer when its expiration period is reached.
B
Retrieval supplies current, inspectable evidence without repeatedly updating model parameters. Weekly fine-tuning with separately retrieved titles does not establish that the references support the generated claims. Daily summaries compress away procedural detail. Semantic caching allows stale answers until expiration and does not tie freshness to source changes.
Foundation Model Integration, Data Management, and Compliance
1
Select One
.png)
.svg)
