Free AWS Certified Generative AI Developer - Professional (AIP-C01) Sample Questions

Correct answer: A — Put an AWS Lambda function behind Amazon API Gateway to abstract the invocation, and source the active model ID from AWS AppConfig so it can change without a code deploy.

• A. Fronting the invocation logic with Lambda and API Gateway, and externalizing the model choice to AppConfig, lets the routing and model ID change through a configuration deployment rather than an application code deploy.
• B. Relying on manual review to catch every hardcoded reference does not remove the underlying coupling and leaves the process error-prone and still tied to a code change.
• C. Environment variables set at build time still require rebuilding and redeploying the container image for every model change, which is the exact friction the team wants removed.
• D. Bundling the model ID inside the deployment package still requires cutting and shipping a new release for every change, so it does not decouple configuration from code deploys.

Correct answer: A — Configure the knowledge base with hierarchical chunking so parent and child chunk relationships are preserved, letting retrieval return a child chunk and expand to its parent when more context is needed.

• A. Hierarchical chunking is designed exactly for this case: it indexes smaller child chunks for precise semantic matching while retaining a link to the larger parent chunk, so retrieval can return the focused answer and still surface surrounding context on demand.
• B. Large fixed-size chunks trade away retrieval precision because a single chunk mixes multiple topics; the model gets more text per hit but loses the ability to zero in on the specific passage that actually answers the question.
• C. Splitting by business unit addresses organizational separation but does nothing for the parent-child depth problem within a single unit's manuals, and juggling five knowledge bases adds retrieval-orchestration complexity the scenario does not call for.
• D. Embedding whole documents produces very coarse similarity matches and forces the FM to consume large amounts of irrelevant text from the context window, which is inefficient and does not target the sub-section the user actually needs.

Correct answer: A — MongoDB Atlas requires manually configuring filters directly in its own vector index settings before metadata filtering will work, since it is not enabled automatically by default.

• A. MongoDB Atlas is documented as requiring the customer to manually configure filters within their own Atlas vector index settings, since metadata filtering does not work out of the box the way it does for some other supported stores.
• B. OpenSearch requires custom metadata fields intended for filtering to be defined with a keyword type or a text field with a keyword subfield; fields mapped without that structure fail filtering queries with a rewrite error rather than working automatically.
• C. Aurora PostgreSQL benefits significantly from a GIN index on the jsonb metadata column for filter performance; relying on sequential scans without an index does not scale and is not how the documented setup recommends configuring filtering.
• D. Amazon S3 Vectors lets metadata be configured as either filterable or non-filterable at index creation time, so treating every key as filterable by default with no option otherwise misstates how S3 Vectors metadata is actually configured.

Correct answer: A — Scoping the problem to determine whether generative AI is an appropriate way to address the stated business need at all.

• A. Defining what success looks like with stakeholders before any model has been chosen or any prompt written is the scoping phase, where the team determines whether generative AI is even the right approach to the business need.
• B. Model selection happens after scoping has established the requirements a model needs to satisfy, but this team has not yet reached that point since no requirements have been agreed upon yet.
• C. Customization work like prompting or fine-tuning happens after a model has already been selected, which has not occurred yet in a meeting held before any model choice or prompt has been discussed.
• D. Iterating on a released capability requires the feature to already be in production collecting usage data, which is far downstream of an initial stakeholder meeting that precedes model selection entirely.

Correct answer: A — Send the image bytes and the description text together in one request to a multimodal foundation model on Amazon Bedrock that accepts combined image and text content in a single prompt.

• A. A multimodal Bedrock model can accept image and text content blocks in the same request, letting it reason jointly over the visual and textual signal in a single inference call.
• B. Rekognition labels only capture generic object tags, discarding the visual detail a multimodal model would otherwise see directly, and the text-only model never receives the image itself.
• C. Storing the image as base64 metadata does not feed it into the model at all, so the title-generation call only ever sees the seller description text.
• D. Textract extracts printed or handwritten text from an image but does not capture non-text visual attributes like color or style, and it still requires a separate text-only inference call.

Correct answer: A — Append each turn as a message object with a `role` of `user` or `assistant` and a `content` array holding the turn's text or media blocks, passed to the Converse API.

• A. The Converse API standardizes conversation state as a list of messages with alternating user/assistant roles and content blocks, working consistently across Bedrock models without a model-specific template.
• B. Manually concatenating turns into a labeled text blob abandons the Converse API's structured message format and reintroduces the model-specific prompt engineering the Converse API is meant to avoid.
• C. Foundation models cannot query external databases mid-inference on their own; the application must retrieve and include the actual turn content in the request.
• D. Foundation models have no memory of a specific ongoing conversation from training data; omitting prior turns means the model loses all context from earlier in the chat.

Correct answers: A, B — Adding a `guardrailConfiguration` field with the guardrail identifier and version to a flow's prompt node so both the input prompt and the generated response are evaluated against the guardrail.; Calling the `ApplyGuardrail` API directly against user input text before it is ever sent to a foundation model, to evaluate the input independently of any model invocation.

• A. A prompt node's guardrail configuration is a documented, supported way to have Bedrock evaluate both the prompt input and the model's completion against the specified guardrail identifier and version as part of flow execution.
• B. The `ApplyGuardrail` API is specifically designed to evaluate arbitrary input or output text against a guardrail independently of invoking a foundation model, making this a valid way to screen user input before any model call.
• C. An S3 storage node's configuration only specifies the target bucket for storing flow data and has no guardrail parameter or built-in mechanism to scan stored files after they are written.
• D. A Lambda function node's configuration only specifies the function ARN to invoke and has no guardrail field, so content filtering on a Lambda's return value is not something Bedrock applies automatically through that node.
• E. A Lex node's configuration specifies the bot alias and locale for intent recognition and has no guardrail parameter, so there is no built-in mechanism for a guardrail to filter a predicted intent through that node type.

Correct answers: A, B — Apply a Bedrock Guardrail with contextual grounding checks configured on the prompt or knowledge base node so ungrounded claims in the generated answer are automatically flagged.; Add an explicit instruction in the prompt directing the model to answer strictly from the supplied trial excerpts and to state clearly when a claim cannot be supported by them.

• A. Contextual grounding checks are the Guardrails feature specifically designed to detect when a generated response introduces information not supported by the retrieved source, which is exactly the automated flagging mechanism compliance is asking for.
• B. An explicit instruction constraining the model to the supplied excerpts and requiring it to acknowledge unsupported claims directly reduces the model's tendency to fabricate content and works alongside the automated grounding check as a complementary safeguard.
• C. A larger output token limit only changes how much text the model is allowed to generate and has no mechanism for detecting or flagging whether a claim in that text is actually grounded in the retrieved excerpts.
• D. Broadening retrieval to unfiltered external literature moves further away from the specific trial reports under review and increases the risk of introducing ungrounded claims rather than helping the assistant stay grounded in the intended source.
• E. Removing summarization entirely would return only raw excerpts without a generated answer, which eliminates the fabrication risk by eliminating the feature itself rather than providing the requested flagging mechanism for generated summaries.

Correct answer: A — Increase the buffer size so more surrounding sentences are combined before embedding, and raise the breakpoint percentile threshold so only clearly dissimilar sentences trigger a new chunk.

• A. A larger buffer size gives each sentence more surrounding context before it is embedded, reducing false splits from momentary wording shifts, while a higher breakpoint percentile threshold requires sentences to be more clearly dissimilar before a boundary is drawn, which together curb both merged and over-split chunks.
• B. Shrinking the buffer size removes the surrounding context that helps distinguish a genuine topic shift from ordinary sentence variation, and a lower breakpoint threshold makes boundaries trigger too easily, worsening the over-splitting problem rather than fixing it.
• C. The maximum tokens parameter only caps chunk size after boundaries are already chosen from sentence dissimilarity, so changing it alone does not address why unrelated paragraphs are merging or a single argument is being split.
• D. Semantic chunking depends on a foundation model to compute sentence dissimilarity by design, and replacing that mechanism with fixed-percentage overlap reverts to a token-based approach that cannot align boundaries with topic shifts.
• E. Pushing the breakpoint threshold to its lowest setting combined with a very small token cap produces excessively fragmented chunks, which trades the merging problem for aggressive over-splitting instead of balancing the two.

Correct answer: A — MCP is a standardized protocol for how agents discover and invoke tools, such as a vector search tool, not a vector database or embedding storage mechanism itself.

• A. MCP is documented as a standardized way for agents to discover and invoke tools, such as a vector search tool exposed behind a gateway, and it does not itself store or index vectors, so it sits alongside a vector store rather than replacing one.
• B. MCP does not physically store or index vector embeddings; that responsibility remains with a vector store such as OpenSearch or Aurora pgvector, so describing MCP as a vector indexing engine misstates its role.
• C. MCP is not an embeddings model and does not convert text into vectors; that conversion is performed by a separate embeddings model such as Titan Text Embeddings, which MCP has no role in replacing.
• D. MCP has nothing to do with how documents are split into chunks during ingestion; chunking strategies operate entirely upstream of, and independently from, the tool-invocation protocol MCP defines.

Correct answer: A — Identity federation issues short-term credentials tied to an existing corporate identity provider, avoiding the burden of provisioning, rotating, and deprovisioning an IAM user per employee.

• A. Federated sign-in through an existing identity provider issues temporary credentials scoped to a session, which removes the need to manually create, rotate, and delete a long-lived IAM user for every one of thousands of employees.
• B. AWS access keys for IAM users are not rotated automatically by AWS; rotation is an operational task the account owner must schedule and perform, so this claim misstates how IAM user credentials are managed.
• C. Long-lived IAM user credentials and short-lived federated session credentials differ meaningfully in exposure window and rotation burden, so the two approaches are not equivalent from a security standpoint.
• D. Bedrock accepts standard AWS IAM authentication, including temporary credentials obtained via federation or assumed roles, so there is no restriction forcing the use of individual long-lived IAM users.

Correct answer: A — The gateway records per-request metrics such as latency, token usage, and error rate for every calling application, and enforces per-application rate limits from that recorded usage.

• A. Recording per-request metrics like latency, token usage, and error rate, and using that data to enforce rate limits, is a direct observability and control mechanism, giving the platform team visibility and enforcement power across every consuming application.
• B. Automatically rewriting prompt text without being asked is a content-manipulation behavior, not an observability or control mechanism, and it risks changing application behavior in ways the calling team never requested.
• C. Storing payloads in a publicly readable, unauthenticated location is a data-exposure risk rather than a control mechanism, and it contradicts the access-control goals a gateway is meant to help enforce.
• D. Removing an application's ability to specify a model takes away a legitimate integration choice rather than adding observability or control, and most gateway designs still let the caller request a specific approved model.

Correct answer: A — They must purchase Amazon Bedrock Provisioned Throughput for the customized model, because Bedrock does not support on-demand invocation for any customized model.

• A. Bedrock requires a Provisioned Throughput purchase before a customized model can be invoked at all — on-demand, pay-per-token invocation is only available for unmodified base foundation models.
• B. Bedrock does host customized models directly; moving to a separate SageMaker AI endpoint is unnecessary and ignores that Bedrock's own customization and Provisioned Throughput features exist for this purpose.
• C. The blocking requirement is Provisioned Throughput, not a token-allowance quota — a quota increase for on-demand throughput would not unlock invocation of a customized model since on-demand access simply is not offered for it.
• D. Agent action groups are for connecting a model to external tools and APIs; they are unrelated to whether a customized model can be invoked, and they do not substitute for the Provisioned Throughput requirement.

Correct answer: A — Serve the chat assistant with a low-latency real-time endpoint kept warm for interactive traffic, score the batch job through batch or asynchronous inference, and serve image captioning through on-demand invocation billed only per request.

• A. A warm real-time endpoint gives the chat assistant the low, consistent latency it needs; batch or asynchronous inference fits document scoring that no user is waiting on; and per-request on-demand invocation avoids any charge while the captioning feature sits idle.
• B. Routing the chat assistant through overnight batch inference would make interactive users wait until the next batch run, and reserving a real-time endpoint continuously for document scoring pays for always-on capacity a background job does not need.
• C. Unpredictable on-demand invocation with no warm capacity risks added latency for a sub-second chat requirement, and sizing a continuously running real-time endpoint for the batch job's peak scoring rate pays for idle capacity between runs.
• D. A single reserved-capacity commitment sized for interactive chat traffic bills continuously regardless of the batch job's and captioning feature's very different, mostly idle usage patterns, which wastes the discount Provisioned Throughput offers steady workloads.

Correct answer: A — Define the tool with a standardized function schema specifying required and optional parameter types, and implement input validation and a consistent, structured response format in the backing Lambda function.

• A. A standardized function schema with typed required and optional parameters, paired with input validation and a consistent structured response, gives any orchestrating model a reliable contract for calling and interpreting the tool.
• B. Relying on free-form prose without a structured schema leaves the expected order ID format and optional region code ambiguous, which increases the chance of malformed calls compared to an explicit schema.
• C. Skipping validation in the Lambda function pushes error handling onto the downstream service and exposes its raw error format to the agent, which is a less controlled and less predictable failure path than validating first.
• D. Passing through an unstable, changing response shape forces every orchestrating model to adapt its parsing individually, which undermines the consistent, predictable tool behavior reliable operation requires.

Correct answer: A — Amazon Bedrock Agents Classic is closed to new customers, and Amazon Bedrock AgentCore is the offering AWS points new builds toward for similar capabilities plus added observability and gateway components.

• A. Bedrock Agents Classic is closed to new customers, and AWS documentation directs new builds toward AgentCore for similar orchestration capabilities along with added observability, identity, and gateway functionality.
• B. This reverses the actual status: Agents Classic is the one closed to new customers, and AgentCore is the actively developed offering AWS points new builds toward, not an unrelated deprecated preview.
• C. AgentCore adds observability, identity, and gateway components beyond what Classic provided, so describing it as an identical rebrand misrepresents the expanded feature set new builds gain by choosing it.
• D. Agents Classic is no longer open to new customers, so the two offerings are not both currently available to a new build in the same way, and the choice is not merely a naming preference.

Correct answer: A — Author an OpenAPI specification that defines the request and response schemas, parameters, and error codes up front, and generate client SDKs and server stubs from that shared contract.

• A. An OpenAPI specification is a machine-readable contract that defines schemas, parameters, and errors before implementation begins, which is exactly what API-first development uses to keep multiple independent clients aligned and to generate consistent SDKs and stubs.
• B. Reading implementation code as the source of truth couples every consuming team to internal details that can change at any time, and it provides no agreed contract to validate requests and responses against.
• C. A prose wiki description is not machine-validated, so it is easy for each team's interpretation to drift from the others, producing inconsistent clients that a formal schema would have caught.
• D. Letting the first integration define the contract by trial and error skips the up-front agreement step entirely and risks baking an accidental, undocumented shape into every later client that copies it.

Correct answer: A — Multi-agent coordination and delegation maps to AWS Agent Squad, an AWS-native framework purpose-built for routing tasks between specialized agents.

• A. AWS Agent Squad is an AWS-native orchestration framework built specifically to route tasks and manage collaboration between multiple specialized agents, which matches the coordinator/worker delegation pattern described.
• B. CloudWatch Synthetics runs scripted health-check canaries against endpoints on a schedule; it has no role in coordinating task delegation between cooperating AI agents.
• C. AWS Config tracks and evaluates the configuration state of AWS resources for compliance purposes, which is unrelated to orchestrating collaboration between AI agents.
• D. EventBridge Scheduler invokes a target on a time-based schedule; it does not provide the task-routing or delegation logic that a multi-agent coordination framework needs.

Correct answers: A, E — API Gateway's HTTP APIs (the newer, lighter-weight API type) do not support chunked transfer encoding, so relying on chunked delivery through that API type will fail even if the backend produces a chunked stream.; API Gateway strips the `Transfer-Encoding` header on affected API types regardless of what the backend integration sends, so the client cannot rely on that header surviving the round trip.

• A. API Gateway's newer HTTP API type does not support chunked transfer encoding, so a design that depends on chunked delivery must avoid that API type and use a different integration path instead.
• B. A standard Lambda proxy integration returns one complete response object to API Gateway, so the function itself must finish assembling output before anything is returned, constraining how incremental the delivery can actually be through that integration style.
• C. Chunked transfer encoding is a one-way HTTP response mechanism and has no relationship to WebSocket connections; it does not convert a REST integration into a bidirectional channel.
• D. Foundation models return their output through the Bedrock runtime's own response format, such as the streaming event structure; nothing about that format is inherently chunked-transfer-encoded HTTP, so backend formatting work is still required.
• E. API Gateway is known to drop or strip the `Transfer-Encoding` header on the API types where chunked encoding is unsupported, so a design cannot depend on that header reaching the client intact.
• F. Chunked transfer encoding is a standard HTTP mechanism that works over ordinary internet connections; it has no dependency on client and backend sharing a VPC.

Correct answer: A — A contextual grounding check in a Bedrock guardrail, configured with a threshold that compares the generated answer against the retrieved source passages and blocks responses that are not sufficiently grounded.

• A. A contextual grounding check specifically scores whether a generated answer is factually supported by and relevant to the retrieved source passages, and can block responses that fall below a configured grounding threshold, which directly targets ungrounded hallucinated claims.
• B. Retrieving more passages can improve the material available to the model but does not itself check whether the eventual generated answer stayed faithful to that material, so ungrounded claims could still pass through unchecked.
• C. A word filter blocks or allows specific literal terms and phrases; it cannot evaluate whether a factual claim in an answer is actually supported by the retrieved documents, so it does not address hallucinated content that uses none of the listed words.
• D. An ingestion failure alarm detects when documents fail to load into the knowledge base, which is a data-pipeline health signal and does not evaluate whether an individual generated answer is grounded in retrieved content.

Correct answer: A — With the Standard tier, content filter and denied topics detection extends into code elements such as comments, variable names, and string literals, so intent hidden in identifiers is still evaluated.

• A. The Standard safeguard tier explicitly extends content filter and denied topics detection into code elements including comments, variable names, and string literals, which is the capability that specifically prevents an attacker from hiding harmful intent inside code identifiers to evade detection.
• B. The sensitive information policy is built to detect structured PII entities like emails and phone numbers, not to recognize harmful intent expressed through descriptive docstrings or variable naming, so it would not reliably catch this framing.
• C. An exact-match word filter only catches the literal phrase it lists, so an attacker who rephrases the same intent with different wording in the docstring or identifiers would evade it entirely, unlike a category-based detection extended into code.
• D. The scenario specifically distinguishes tier behavior, and the Classic tier does not extend content and topic detection into code elements the way the Standard tier does, so treating them as equivalent misses the actual difference relevant here.

Correct answer: A — An Amazon SageMaker Model Card documenting intended use, known limitations, and evaluation results in one record.

• A. A model card is purpose-built to centralize intended use, risk rating, limitations, and evaluation results in one governed record, which is exactly the single documentation artifact risk and compliance stakeholders can review.
• B. A guardrail enforces safety and policy rules at inference time; it does not document intended use, known limitations, or evaluation results as a governance artifact for stakeholder review.
• C. An operational dashboard reports runtime metrics like latency and token usage, which is useful for monitoring but does not capture the model's intended use, limitations, or evaluation findings.
• D. An IAM policy controls access permissions to the model endpoint; it has no role in documenting the model's use cases, limitations, or evaluation results.

Correct answer: A — A contextual grounding check that compares the generated response against the retrieved source and flags unsupported claims.

• A. A contextual grounding check exists precisely to compare a generated response against the retrieved source and flag or block content that is not grounded in it, which is what catches an invented interaction claim.
• B. A denied topics filter blocking all drug-interaction mentions would also block accurate, correctly grounded warnings, which is not targeted at distinguishing an invented claim from a supported one.
• C. A word filter matches specific listed terms; it has no mechanism for comparing a generated statement against retrieved source material to determine whether the claim is actually supported.
• D. The misconduct content category targets a specific class of harmful content unrelated to factual grounding, so raising its strength does nothing to detect whether a medical claim matches the retrieved source.

Correct answers: A, B, C — Table-level and column-level permissions that let administrators grant a department read access to specific tables while excluding sensitive columns such as payment fields.; Data filters that provide row and cell-level security so the same underlying table returns different rows or masked cells depending on the requesting principal's grant.; LF-Tags that let administrators tag tables and columns once and grant access by tag expression, so department access scales without per-table permission edits.

• A. Column-level permissions let an administrator grant a department access to a table while excluding specific columns such as payment fields, which is exactly the fine-grained exclusion the requirement describes. This works on the single centralized copy of the data, so no duplication is needed.
• B. Data filters apply row and cell-level rules so different departments querying the same physical table see only their department's rows or masked versions of restricted cells. This satisfies both the department scoping and payment-field protection without copying data.
• C. LF-Tags let administrators tag resources once and grant access through tag expressions, which scales department-based access as new tables are added without re-granting permissions table by table. This directly supports centralized, non-duplicated access management.
• D. Copying tables into per-department accounts creates physical duplicates of the same data, which directly contradicts the requirement to avoid duplicating data per department. It also multiplies the operational burden of keeping copies synchronized.
• E. Query result caching only affects performance by reducing repeated scans; it has no access control mechanism and does nothing to restrict which tables, rows, or columns a department can see.
• F. Blueprint workflows automate data ingestion from source systems into the catalog, which addresses how data arrives rather than how access to that data is restricted once it is there. This does not provide any department- or column-level access control.

Correct answer: A — Apply deterministic tokenization at the application layer before sending text to the model, replacing each account number and name with a consistent token per customer for the session.

• A. Deterministic tokenization maps the same account number or name to the same token every time within a session, so the model never sees the raw value but can still track that later messages refer to the same customer, which follow-up questions depend on.
• B. Generating a fresh random string for every message breaks the mapping between a customer and their masked identifier from one message to the next, so the model cannot recognize that later messages refer to the same account.
• C. The model's context window retains whatever raw text is sent to it during a session; it does not automatically strip or forget sensitive values, so raw account numbers and names would still reach the model with no masking applied at all.
• D. Contextual grounding checks evaluate whether a response is factually supported by retrieved source content; they are designed to catch hallucinations, not to mask or replace sensitive values before the model ever sees them.

Correct answers: A, B, E — A requirement that every deployed model have a completed model card documenting its intended use, evaluation results, and risk rating on file.; A cross-functional review board that approves new GenAI use cases against the policy framework and regulatory requirements before production deployment begins.; A central policy framework defining minimum required controls, like guardrail categories and logging destinations, every application must satisfy at launch.

• A. Requiring a completed model card on file for every deployed model gives the program a consistent, auditable artifact per model, aligning organizational oversight with the same documentation standard across every business unit.
• B. A cross-functional review board that checks each new use case against the policy framework before production is the structural gate that prevents inconsistent, ungoverned launches while still letting teams ship once they pass review.
• C. A policy with no review cadence goes stale as regulations and business needs change, so it fails to provide the consistent, current oversight a governance program is meant to maintain over time.
• D. Centralizing every prompt and guardrail configuration change through email requests to one team creates a bottleneck that directly contradicts the risk officer's requirement to keep teams able to ship, rather than establishing scalable oversight.
• E. A central policy framework setting minimum required controls gives every business unit a shared, consistent baseline to build against, which is the mechanism that keeps oversight consistent across units that otherwise choose their own models and controls independently.

Correct answer: B — No, because CloudTrail logs that InvokeModel ran and by whom, not the response text; reconstructing it needs model invocation logging.

• A. Bedrock does not forward model response bodies into CloudTrail events; response content capture is the purpose of the separate model invocation logging feature, which writes to CloudWatch Logs or S3, not to CloudTrail.
• B. CloudTrail is an API activity log: it records that a principal called InvokeModel, on which model, and when, but the event's responseElements field does not carry the model's response body, so reconstructing the exact returned text requires Bedrock model invocation logging instead, delivered to CloudWatch Logs or S3.
• C. CloudTrail records both console actions and programmatic SDK or CLI API calls; it is not limited to console sign-in events, so this understates what CloudTrail actually covers even though it correctly concludes CloudTrail cannot reconstruct the response text.
• D. CloudTrail's own published example log entry for the InvokeModel action shows the responseElements field as null: the event records that the call happened and its request parameters, not the model's output text, so there is no response text to view in Event history even though InvokeModel is logged by default as a management event.

Correct answers: A, B — Measure the peak input and output token rate the workload actually needs per minute during the busiest observed traffic period, and size the Model Unit count to cover that peak rather than the daily average.; Review historical or projected traffic patterns to choose a commitment term, no commitment, one month, or six months, matching how confident the team is in sustained demand versus flexibility.

• A. Provisioned Throughput capacity is defined by how many input and output tokens per minute a Model Unit can sustain, so sizing to the actual peak rate is what prevents throttling during the busiest periods without over-provisioning for the rest of the day.
• B. The commitment term directly trades flexibility for a lower hourly rate, so matching the term to how confident the team is in sustained demand avoids either paying a premium for unnecessary flexibility or locking into capacity longer than justified.
• C. Provisioned Throughput can be modified or deleted, subject to the chosen commitment term, and provisioning the account maximum regardless of measured traffic wastes money on capacity the workload does not use.
• D. A Model Unit delivers a specific, bounded throughput level, not unlimited throughput, so a single unit sized without reference to actual measured traffic will throttle a workload whose real demand exceeds that unit's capacity.
• E. On-demand and Provisioned Throughput use different pricing models entirely, hourly commitment versus per-token, so the number of Model Units purchased directly determines fixed hourly cost independent of how many tokens are actually processed.

Correct answer: D — Continuously monitor OpenSearch Serverless CloudWatch latency and request-rate metrics, and schedule automated index optimization routines such as force-merge to consolidate segments.

• A. A relational database with a JSON column has no native approximate nearest neighbor vector index at all, so it is not a comparable vector store and does not address the operational monitoring gap the scenario describes.
• B. Larger chunks reduce vector count but change retrieval granularity and answer quality, and chunk size has no direct relationship to segment fragmentation accumulated from continuous document ingestion over months.
• C. Re-embedding with a newer model changes vector representations and could even require reindexing, but it does not address segment fragmentation from ongoing ingestion, which is the mechanism behind steadily growing search latency with stable volume.
• D. Tracking OpenSearch Serverless's own latency and request-rate metrics over time would show the gradual trend directly, and scheduled index optimization routines address the underlying cause of segment fragmentation from continuous ingestion that typically drives this kind of slow latency creep.

Correct answer: A — Run an A/B test that splits live traffic between the current and new templates and compares first-reply resolution rate between the two groups over the same period.

• A. An A/B test that runs both templates concurrently against comparable live traffic isolates the effect of the prompt change itself, giving a direct, evidence-based comparison of first-reply resolution rate between the two versions.
• B. Raising temperature changes how varied the wording sounds but has no bearing on whether the underlying prompt template actually resolves more customer questions, and it does not produce a comparison against the old template at all.
• C. Prompt caching can reduce response latency, but faster responses are not evidence of better resolution quality, and this approach never compares the two templates against each other.
• D. Comparing the new template's live performance against a prior quarter's historical average confounds seasonal and traffic differences with the prompt change, so it does not isolate whether the new template caused the difference.

Correct answers: A, B — Track prompt and completion token volume and invocation counts against throughput quotas, since GenAI cost is driven by token throughput, not server CPU.; Track queue depth and the number of concurrent in-flight Bedrock requests against the account's configured concurrency limits during traffic spikes.

• A. Because the application servers are just thin callers to Bedrock, the real constraint during a spike is token throughput against the account's per-minute quotas, so surfacing prompt/completion token volume and invocation-rate metrics reveals the actual bottleneck that CPU and memory cannot show.
• B. Queue depth and concurrent in-flight request counts reveal exactly when the fleet is approaching its effective concurrency ceiling, giving a second leading indicator that CPU and memory metrics cannot surface.
• C. Disk I/O on the calling servers is unrelated to Bedrock's model-side throughput limits; the application servers making lightweight API calls are unlikely to be storage-bound, so this metric would not explain the timeouts.
• D. Raising the CPU and memory thresholds would delay scaling further and does not address that CPU and memory were never the constraint, since they were already low when the timeouts occurred.
• E. A fixed daily schedule adds capacity independent of actual traffic or token-throughput conditions, so it would not target the load spikes that are the actual cause of the timeouts.

Correct answer: A — Route the rated responses into an annotation workflow that groups low-rated responses by failure pattern, then feed the insights into targeted changes that are re-evaluated before rollout.

• A. An annotation workflow that clusters low-rated responses by the specific way they failed converts raw ratings into actionable patterns, letting the team make targeted changes and re-evaluate them, which closes the feedback loop the scenario asks for.
• B. Deleting negative ratings hides the underlying quality problem rather than fixing it, and it removes exactly the signal that would drive a genuine improvement in the prompt or configuration.
• C. A raw thumbs-up count with no analysis of the negative responses provides a vanity metric but no insight into what is going wrong, so it cannot drive targeted prompt or configuration changes.
• D. Structuring only the positive feedback while leaving negative feedback unstructured discards the responses most useful for identifying failure patterns, since thumbs-down ratings are the ones that need categorization to act on.

Correct answer: A — Run automated quality gates and regression tests against known cases first, then proceed to broader human or LLM-as-a-judge review and a staged canary rollout only after the automated checks pass.

• A. Running automated quality gates and regression tests first catches known-failure patterns cheaply and quickly, and only promoting a change to broader review and staged rollout after passing those checks forms the systematic order a mature QA process follows.
• B. Rolling out to full production before any automated regression testing removes the early, low-cost detection step and turns customer complaints into the first line of defense, which is the reverse of a systematic quality assurance process.
• C. A single engineer's personal impression is not a repeatable or systematic check, and skipping both automated testing and staged rollout removes the structured gates a quality assurance process depends on.
• D. Rolling out to all traffic before running any regression tests defeats the purpose of a canary rollout, which is meant to limit exposure while the change is still being validated, not to be run at full scale first.

Correct answer: A — The request body does not conform to the target model's expected schema or parameter constraints, so the developer should compare the request's JSON body against the model's documented schema.

• A. A ValidationException signals a malformed request such as wrong parameter names, types, or out-of-range values for that specific model, so comparing the payload against the documented schema is the right first check.
• B. A billing suspension produces an account-level access failure affecting all requests, not a per-request ValidationException tied to a specific malformed body, so it would not explain selective failures.
• C. Regional service maintenance typically surfaces as throttling or service-unavailable errors across most requests, not a validation error scoped to a particular request's content.
• D. A missing invoke permission produces an AccessDeniedException rather than a ValidationException, so a policy gap does not explain this specific error code.

Correct answers: A, B, C — Isolate which specific instruction changed between the two template versions and test that change independently against the same evaluation set to confirm it is the source of the accuracy regression.; Re-run the lower-scoring template version several times against the same inputs to check whether the inconsistency is inherent to the prompt wording or attributable to sampling variance at the current temperature setting.; Iterate on the instruction wording that caused the regression, for example by adding explicit grounding or verification steps, then re-score the revised version against the same evaluation set before promoting it.

• A. Isolating the specific instruction change and testing it independently is the systematic way to confirm which wording caused the accuracy regression before touching anything else.
• B. Checking whether inconsistency persists across repeated runs distinguishes an inherent prompt problem from ordinary sampling variance, which changes what fix is actually needed.
• C. Iteratively refining the identified problem wording and re-scoring against the same evaluation set is the core loop of a systematic, repeatable refinement workflow.
• D. Promoting a template with a confirmed factual accuracy regression in a customer-support context without addressing the cause ignores the evaluation results rather than acting on them.
• E. A single template's consistency issue does not justify a blanket, permanent temperature change across every other template in the application; that generalizes one finding far beyond its evidence.