Topic 12 Practice Test 2 covers GenAI Cost Optimization, Caching, Resource Efficiency, and Performance Engineering in Domain 4 (Tasks 4.1-4.2). For broader exam preparation, review the AIP-C01 Exam Dumps. Every option includes focused technical reasoning tied to the tested decision.
Question 1
An application team is standardizing data before foundation-model inference. A support assistant sends the full conversation and 30 retrieved passages on every turn even though only a few are relevant. Which implementation provides the required behavior? Choose ONE.
- context pruning by relevance
- retrieval top-k reduction after quality testing
- conditional context inclusion
- maximum response-token limit
Correct Answer: A
Correct Answer
Answer A is correct because context pruning by relevance is correct. It remove low-value history and retrieved passages while preserving information needed for the current request. This matters because prompt context has grown over time and includes material unrelated to the current question. That directly addresses the scenario.
Incorrect Answers
Answer B is incorrect because retrieval top-k reduction after quality testing has value because it can send fewer high-value retrieved chunks when evaluation shows additional chunks do not improve answer quality. Its proper fit is when RAG context is expensive because too many marginal passages are inserted into every prompt. This question requires prompt context has grown over time and includes material unrelated to the current question.
Answer C is incorrect because conditional context inclusion can help a pipeline attach expensive reference context only for requests that actually need it, especially when all requests currently carry a large reference bundle even when simple questions can be answered without it. The present requirement is prompt context has grown over time and includes material unrelated to the current question, so it is not the best answer.
Answer D is incorrect because maximum response-token limit would improve another stage by enabling it to cap generation length so routine responses cannot consume an excessive number of completion tokens. It fits when the business requires concise answers and wants a hard upper bound on completion cost. The stated problem is prompt context has grown over time and includes material unrelated to the current question.
Question 2
An operations team is improving the reliability of an FM input path. A production prompt repeats the same policy language in multiple sections and consumes thousands of avoidable tokens. Which change should the team make? Choose ONE.
- prompt compression
- retrieval top-k reduction after quality testing
- structured concise system prompt
- token counting before invocation
Correct Answer: A
Correct Answer
Answer A is correct because Choose prompt compression. Its role is to rewrite verbose instructions and repeated context into a shorter representation that preserves essential meaning. The key requirement is that prompt tokens are dominated by redundant wording rather than necessary evidence.
Incorrect Answers
Answer B is incorrect because retrieval top-k reduction after quality testing belongs in workflows where RAG context is expensive because too many marginal passages are inserted into every prompt because it can send fewer high-value retrieved chunks when evaluation shows additional chunks do not improve answer quality. This question instead asks for prompt tokens are dominated by redundant wording rather than necessary evidence.
Answer C is incorrect because structured concise system prompt is useful when system instructions are unnecessarily verbose and repeated on each request and can remove duplicated prose and express stable constraints in a compact structured instruction block. The described workload requires prompt tokens are dominated by redundant wording rather than necessary evidence.
Answer D is incorrect because token counting before invocation targets a different issue by helping to estimate prompt and expected completion tokens before sending the request so oversized or unexpectedly expensive calls can be controlled. It fits where the application needs predictable cost and context-window use before FM inference. The deciding condition is prompt tokens are dominated by redundant wording rather than necessary evidence.
Question 3
A developer is hardening an FM-consumption workflow. A customer conversation can last hundreds of turns and eventually approaches the context limit. Which approach most directly addresses the stated constraint? Choose ONE.
- maximum response-token limit
- conditional context inclusion
- structured concise system prompt
- conversation summarization for long history
Correct Answer: D
Correct Answer
Answer D is correct because conversation summarization for long history is appropriate because it will replace older detailed turns with a compact maintained summary while retaining recent conversation verbatim. The requirement centers on a situation where long-running chats need continuity without resending every historical token.
Incorrect Answers
Answer A is incorrect because maximum response-token limit is designed to cap generation length so routine responses cannot consume an excessive number of completion tokens. That makes sense when the business requires concise answers and wants a hard upper bound on completion cost. The question is centered on long-running chats need continuity without resending every historical token.
Answer B is incorrect because conditional context inclusion can attach expensive reference context only for requests that actually need it. Its best fit is where all requests currently carry a large reference bundle even when simple questions can be answered without it. The current case instead needs long-running chats need continuity without resending every historical token.
Answer C is incorrect because structured concise system prompt helps when system instructions are unnecessarily verbose and repeated on each request by allowing the pipeline to remove duplicated prose and express stable constraints in a compact structured instruction block. The described issue is long-running chats need continuity without resending every historical token.
Question 4
A GenAI workload is moving from prototype to production. An extraction service needs exactly five fields, but the model currently returns a long narrative around each result. Which option best meets the operational requirement? Choose ONE.
- response schema with concise fields
- token counting before invocation
- conditional context inclusion
- retrieval top-k reduction after quality testing
Correct Answer: A
Correct Answer
Answer A is correct because The requirement points to response schema with concise fields. It is meant to request a bounded structured output containing only fields the application consumes. The scenario needs an approach where free-form generation includes explanations the downstream system discards.
Incorrect Answers
Answer B is incorrect because Using token counting before invocation would let the team estimate prompt and expected completion tokens before sending the request so oversized or unexpectedly expensive calls can be controlled. That is appropriate when the application needs predictable cost and context-window use before FM inference. The scenario instead calls for free-form generation includes explanations the downstream system discards.
Answer C is incorrect because conditional context inclusion has value because it can attach expensive reference context only for requests that actually need it. Its proper fit is when all requests currently carry a large reference bundle even when simple questions can be answered without it. This question requires free-form generation includes explanations the downstream system discards.
Answer D is incorrect because retrieval top-k reduction after quality testing can help a pipeline send fewer high-value retrieved chunks when evaluation shows additional chunks do not improve answer quality, especially when RAG context is expensive because too many marginal passages are inserted into every prompt. The present requirement is free-form generation includes explanations the downstream system discards, so it is not the best answer.
Question 5
A regulated workflow requires deterministic preprocessing before inference. Monthly Bedrock spend rose sharply, but the team cannot tell which product feature or request type caused the increase. Which solution most directly resolves this issue? Choose ONE.
- maximum response-token limit
- token-usage telemetry by feature
- structured concise system prompt
- token counting before invocation
Correct Answer: B
Correct Answer
Answer B is correct because token-usage telemetry by feature matches the constraint. It can record prompt and completion usage per workflow so the team can identify the largest cost drivers. This is the intended approach when cost reduction requires knowing which application paths consume the most tokens.
Incorrect Answers
Answer A is incorrect because maximum response-token limit would be sound if the goal were the business requires concise answers and wants a hard upper bound on completion cost; it can cap generation length so routine responses cannot consume an excessive number of completion tokens. The requirement here is cost reduction requires knowing which application paths consume the most tokens.
Answer C is incorrect because structured concise system prompt belongs in workflows where system instructions are unnecessarily verbose and repeated on each request because it can remove duplicated prose and express stable constraints in a compact structured instruction block. This question instead asks for cost reduction requires knowing which application paths consume the most tokens.
Answer D is incorrect because token counting before invocation is useful when the application needs predictable cost and context-window use before FM inference and can estimate prompt and expected completion tokens before sending the request so oversized or unexpectedly expensive calls can be controlled. The described workload requires cost reduction requires knowing which application paths consume the most tokens.
Question 6
During a production-readiness review, engineers identify a data-handling issue. A support application handles both simple FAQ classification and difficult multi-document reasoning. Which design is the best fit for this requirement? Choose ONE.
- quality-adjusted price benchmarking
- task-specific model mapping
- batch inference for offline workloads
- complexity-based tiered model routing
Correct Answer: D
Correct Answer
Answer D is correct because Use complexity-based tiered model routing for this design. It will classify requests by difficulty and select a model tier that balances quality and inference price. That aligns with the need that workload complexity varies significantly and a single model tier wastes cost on easy cases.
Incorrect Answers
Answer A is incorrect because quality-adjusted price benchmarking targets a neighboring concern. It can compare candidate models using measured task quality together with per-request cost rather than price alone when the cheapest model may require retries or fail quality thresholds, so selection must consider value. The key requirement here is workload complexity varies significantly and a single model tier wastes cost on easy cases.
Answer B is incorrect because task-specific model mapping is designed to assign different approved models to summarization, extraction, and reasoning based on benchmarked capability. That makes sense when one model is not the best price-performance choice for every task type. The question is centered on workload complexity varies significantly and a single model tier wastes cost on easy cases.
Answer C is incorrect because batch inference for offline workloads can process noninteractive requests in batches rather than paying the latency overhead of individual synchronous calls. Its best fit is where large jobs have no real-time response requirement and can be grouped efficiently. The current case instead needs workload complexity varies significantly and a single model tier wastes cost on easy cases.
Question 7
A cloud engineering team must correct a data-processing weakness. A document classifier is correct on routine inputs with a smaller model but struggles on ambiguous cases. Which implementation best satisfies the requirement? Choose ONE.
- small-model routing for simple requests
- task-specific model mapping
- provisioned throughput for steady high utilization
- fallback escalation to a larger model
Correct Answer: D
Correct Answer
Answer D is correct because fallback escalation to a larger model is the best fit because it can start with an efficient model and escalate only low-confidence or failed cases to a more capable model. The workflow needs a case where most requests are easy but a minority need stronger reasoning.
Incorrect Answers
Answer A is incorrect because small-model routing for simple requests is useful when many requests do not require the capability or price of the most powerful FM because it can send low-complexity tasks to a less expensive capable model while reserving larger models for harder work. The current requirement is most requests are easy but a minority need stronger reasoning, which makes this a mismatch.
Answer B is incorrect because task-specific model mapping addresses another concern by helping to assign different approved models to summarization, extraction, and reasoning based on benchmarked capability. It fits where one model is not the best price-performance choice for every task type. Here the deciding need is most requests are easy but a minority need stronger reasoning.
Answer C is incorrect because Using provisioned throughput for steady high utilization would let the team reserve model throughput when sustained predictable demand can justify committed capacity and reduce capacity uncertainty. That is appropriate when traffic is consistently high and utilization can keep reserved capacity productive. The scenario instead calls for most requests are easy but a minority need stronger reasoning.
Question 8
A solutions architect is reviewing how source data reaches an FM. An internal legal assistant receives only a few requests per hour and is unused overnight. What should be used to meet this requirement? Choose ONE.
- quality-adjusted price benchmarking
- on-demand inference for sporadic traffic
- batch inference for offline workloads
- provisioned throughput for steady high utilization
Correct Answer: B
Correct Answer
Answer B is correct because Implement on-demand inference for sporadic traffic. It is designed to use pay-per-request capacity when traffic is intermittent and does not justify continuously reserved throughput. Here, the important constraint is that request volume is low and bursty with long idle periods.
Incorrect Answers
Answer A is incorrect because quality-adjusted price benchmarking is appropriate where the cheapest model may require retries or fail quality thresholds, so selection must consider value. It achieves that by helping to compare candidate models using measured task quality together with per-request cost rather than price alone. This scenario instead requires request volume is low and bursty with long idle periods.
Answer C is incorrect because batch inference for offline workloads can process noninteractive requests in batches rather than paying the latency overhead of individual synchronous calls, which matters when large jobs have no real-time response requirement and can be grouped efficiently. The present case is driven by request volume is low and bursty with long idle periods.
Answer D is incorrect because provisioned throughput for steady high utilization belongs in workflows where traffic is consistently high and utilization can keep reserved capacity productive because it can reserve model throughput when sustained predictable demand can justify committed capacity and reduce capacity uncertainty. This question instead asks for request volume is low and bursty with long idle periods.
Question 9
A platform team is reviewing a production GenAI data pipeline. Several approved models are available, but responses must score at least 0.92 on the organization’s evaluation set. What is the most appropriate implementation choice? Choose ONE.
- small-model routing for simple requests
- cost-aware route with quality floor
- batch inference for offline workloads
- task-specific model mapping
Correct Answer: B
Correct Answer
Answer B is correct because cost-aware route with quality floor should be used. It can choose the cheapest eligible model only after filtering candidates that meet the required quality threshold. That capability is needed because the optimization objective is cost but quality cannot fall below a contractual minimum.
Incorrect Answers
Answer A is incorrect because small-model routing for simple requests works for cases where many requests do not require the capability or price of the most powerful FM by helping to send low-complexity tasks to a less expensive capable model while reserving larger models for harder work. The present scenario needs the optimization objective is cost but quality cannot fall below a contractual minimum.
Answer C is incorrect because batch inference for offline workloads is designed to process noninteractive requests in batches rather than paying the latency overhead of individual synchronous calls. That makes sense when large jobs have no real-time response requirement and can be grouped efficiently. The question is centered on the optimization objective is cost but quality cannot fall below a contractual minimum.
Answer D is incorrect because task-specific model mapping can assign different approved models to summarization, extraction, and reasoning based on benchmarked capability. Its best fit is where one model is not the best price-performance choice for every task type. The current case instead needs the optimization objective is cost but quality cannot fall below a contractual minimum.
Question 10
An enterprise team is refining the input pipeline for a GenAI application. A product uses Bedrock for live chat during the day and a large offline summarization job at night. What should the developer implement? Choose ONE.
- quality-adjusted price benchmarking
- small-model routing for simple requests
- separate real-time and batch model paths
- provisioned throughput for steady high utilization
Correct Answer: C
Correct Answer
Answer C is correct because separate real-time and batch model paths solves the right problem. It will route interactive and offline workloads to different inference patterns optimized for their latency and cost needs. The case depends on the fact that one architecture currently forces both user-facing and nightly processing through the same synchronous path.
Incorrect Answers
Answer A is incorrect because quality-adjusted price benchmarking can compare candidate models using measured task quality together with per-request cost rather than price alone. It fits when the cheapest model may require retries or fail quality thresholds, so selection must consider value. This scenario instead requires one architecture currently forces both user-facing and nightly processing through the same synchronous path, so it targets a different problem.
Answer B is incorrect because small-model routing for simple requests is useful when many requests do not require the capability or price of the most powerful FM because it can send low-complexity tasks to a less expensive capable model while reserving larger models for harder work. The current requirement is one architecture currently forces both user-facing and nightly processing through the same synchronous path, which makes this a mismatch.
Answer D is incorrect because Using provisioned throughput for steady high utilization would let the team reserve model throughput when sustained predictable demand can justify committed capacity and reduce capacity uncertainty. That is appropriate when traffic is consistently high and utilization can keep reserved capacity productive. The scenario instead calls for one architecture currently forces both user-facing and nightly processing through the same synchronous path.
Question 11
An application team is standardizing data before foundation-model inference. Two services each send 100 requests per minute, but one uses ten times more tokens per request. Which implementation provides the required behavior? Choose ONE.
- separate capacity pools by workload priority
- concurrency control at the caller
- provisioned-throughput utilization monitoring
- token-throughput capacity planning
Correct Answer: D
Correct Answer
Answer D is correct because token-throughput capacity planning is the direct remedy. It can size inference capacity using observed input and output token rates rather than request count alone. The critical scenario condition is that requests vary widely in prompt and completion size, making raw requests-per-second misleading.
Incorrect Answers
Answer A is incorrect because separate capacity pools by workload priority solves a different input problem. It helps to isolate high-priority interactive traffic from background jobs so batch demand cannot consume all available throughput, normally when user-facing requests require predictable performance while offline jobs are flexible. The current workflow needs requests vary widely in prompt and completion size, making raw requests-per-second misleading.
Answer B is incorrect because concurrency control at the caller is appropriate where unbounded parallelism creates throttling and retries that reduce effective throughput. It achieves that by helping to limit simultaneous model invocations to a level the downstream quota and application can sustain. This scenario instead requires requests vary widely in prompt and completion size, making raw requests-per-second misleading.
Answer C is incorrect because provisioned-throughput utilization monitoring would be sound if the goal were the team needs to avoid both idle reserved capacity and chronic capacity pressure; it can track reserved-capacity usage and adjust commitments when utilization remains too low or saturation remains too high. The requirement here is requests vary widely in prompt and completion size, making raw requests-per-second misleading.
Question 12
An operations team is improving the reliability of an FM input path. An SQS-backed inference service alternates between long idle periods and large bursts of pending jobs. Which change should the team make? Choose ONE.
- request batching for compatible work
- provisioned-throughput utilization monitoring
- backpressure with durable queueing
- autoscaling from queue depth and processing rate
Correct Answer: D
Correct Answer
Answer D is correct because Use autoscaling from queue depth and processing rate. It will scale consumers based on backlog and measured token-processing capacity instead of fixed instance counts. The deciding condition is that asynchronous GenAI work accumulates in a queue and demand changes throughout the day. This matches the requirement.
Incorrect Answers
Answer A is incorrect because request batching for compatible work would improve another stage by enabling it to combine compatible small inference items into supported batches to improve throughput and reduce per-request overhead. It fits when many independent small items can be processed together without user-facing latency requirements. The stated problem is asynchronous GenAI work accumulates in a queue and demand changes throughout the day.
Answer B is incorrect because provisioned-throughput utilization monitoring works for cases where the team needs to avoid both idle reserved capacity and chronic capacity pressure by helping to track reserved-capacity usage and adjust commitments when utilization remains too low or saturation remains too high. The present scenario needs asynchronous GenAI work accumulates in a queue and demand changes throughout the day.
Answer C is incorrect because backpressure with durable queueing targets a neighboring concern. It can accept work into a queue and process it at sustainable throughput instead of letting producers overwhelm inference capacity when incoming demand can temporarily exceed available model throughput but requests must not be lost. The key requirement here is asynchronous GenAI work accumulates in a queue and demand changes throughout the day.
Question 13
A developer is hardening an FM-consumption workflow. A report generator produces three independent section summaries before a final aggregation step. Which approach most directly addresses the stated constraint? Choose ONE.
- separate capacity pools by workload priority
- concurrency control at the caller
- parallelization of independent subrequests
- backpressure with durable queueing
Correct Answer: C
Correct Answer
Answer C is correct because parallelization of independent subrequests directly fits. It can run independent model or retrieval operations concurrently when their results do not depend on one another. The scenario is specifically about the case where serial execution adds latency even though several workflow branches have no dependency.
Incorrect Answers
Answer A is incorrect because separate capacity pools by workload priority targets a different issue by helping to isolate high-priority interactive traffic from background jobs so batch demand cannot consume all available throughput. It fits where user-facing requests require predictable performance while offline jobs are flexible. The deciding condition is serial execution adds latency even though several workflow branches have no dependency.
Answer B is incorrect because concurrency control at the caller can limit simultaneous model invocations to a level the downstream quota and application can sustain. It fits when unbounded parallelism creates throttling and retries that reduce effective throughput. This scenario instead requires serial execution adds latency even though several workflow branches have no dependency, so it targets a different problem.
Answer D is incorrect because backpressure with durable queueing addresses another concern by helping to accept work into a queue and process it at sustainable throughput instead of letting producers overwhelm inference capacity. It fits where incoming demand can temporarily exceed available model throughput but requests must not be lost. Here the deciding need is serial execution adds latency even though several workflow branches have no dependency.
Question 14
A GenAI workload is moving from prototype to production. A batch processor is fast at moderate concurrency but becomes slower overall when throttling rises. Which option best meets the operational requirement? Choose ONE.
- request batching for compatible work
- adaptive worker concurrency
- provisioned-throughput utilization monitoring
- separate capacity pools by workload priority
Correct Answer: B
Correct Answer
Answer B is correct because The correct choice is adaptive worker concurrency. It helps to change the number of simultaneous inference workers using observed latency and throttling rather than a permanently high setting. The deciding detail is that optimal parallelism changes as token sizes and service conditions change.
Incorrect Answers
Answer A is incorrect because request batching for compatible work is relevant, but it is intended to combine compatible small inference items into supported batches to improve throughput and reduce per-request overhead when many independent small items can be processed together without user-facing latency requirements. The actual requirement is optimal parallelism changes as token sizes and service conditions change.
Answer C is incorrect because provisioned-throughput utilization monitoring is appropriate where the team needs to avoid both idle reserved capacity and chronic capacity pressure. It achieves that by helping to track reserved-capacity usage and adjust commitments when utilization remains too low or saturation remains too high. This scenario instead requires optimal parallelism changes as token sizes and service conditions change.
Answer D is incorrect because separate capacity pools by workload priority would be sound if the goal were user-facing requests require predictable performance while offline jobs are flexible; it can isolate high-priority interactive traffic from background jobs so batch demand cannot consume all available throughput. The requirement here is optimal parallelism changes as token sizes and service conditions change.
Question 15
A regulated workflow requires deterministic preprocessing before inference. A compliance summarization job must finish by 6 a.m. but currently starts during the evening traffic peak. Which solution most directly resolves this issue? Choose ONE.
- concurrency control at the caller
- request batching for compatible work
- off-peak scheduling for flexible batch jobs
- backpressure with durable queueing
Correct Answer: C
Correct Answer
Answer C is correct because Select off-peak scheduling for flexible batch jobs. Its purpose is to move nonurgent high-volume inference to periods with lower competing demand when business timing permits. The operational requirement is that the workload has a completion deadline but does not need to run during the busiest interactive window.
Incorrect Answers
Answer A is incorrect because concurrency control at the caller can help a pipeline limit simultaneous model invocations to a level the downstream quota and application can sustain, especially when unbounded parallelism creates throttling and retries that reduce effective throughput. The present requirement is the workload has a completion deadline but does not need to run during the busiest interactive window, so it is not the best answer.
Answer B is incorrect because request batching for compatible work would improve another stage by enabling it to combine compatible small inference items into supported batches to improve throughput and reduce per-request overhead. It fits when many independent small items can be processed together without user-facing latency requirements. The stated problem is the workload has a completion deadline but does not need to run during the busiest interactive window.
Answer D is incorrect because backpressure with durable queueing targets a neighboring concern. It can accept work into a queue and process it at sustainable throughput instead of letting producers overwhelm inference capacity when incoming demand can temporarily exceed available model throughput but requests must not be lost. The key requirement here is the workload has a completion deadline but does not need to run during the busiest interactive window.
Question 16
During a production-readiness review, engineers identify a data-handling issue. A backend repeatedly submits the same normalized document and extraction prompt during retry storms. Which design is the best fit for this requirement? Choose ONE.
- prompt caching for repeated long prefixes
- cache key includes authorization context
- deterministic request-hash cache
- hybrid lexical and vector retrieval
Correct Answer: C
Correct Answer
Answer C is correct because The strongest answer is deterministic request-hash cache. It works by helping to cache results by a canonical hash when identical deterministic inputs should always reuse the same output. The requirement is that the workload receives byte-equivalent or canonically equivalent requests with stable expected responses.
Incorrect Answers
Answer A is incorrect because prompt caching for repeated long prefixes is useful when a substantial portion of input tokens is identical across many invocations and can reuse supported cached prompt context when many requests share a large stable instruction or document prefix. The described workload requires the workload receives byte-equivalent or canonically equivalent requests with stable expected responses.
Answer B is incorrect because cache key includes authorization context targets a different issue by helping to include tenant, permissions, or data-scope attributes in the key so cached content is not reused across security boundaries. It fits where semantically identical questions can legitimately produce different answers for different users. The deciding condition is the workload receives byte-equivalent or canonically equivalent requests with stable expected responses.
Answer D is incorrect because hybrid lexical and vector retrieval is useful when pure vector search misses exact product codes while pure keyword search misses paraphrased intent because it can combine keyword and semantic signals when exact identifiers and conceptual meaning are both important. The current requirement is the workload receives byte-equivalent or canonically equivalent requests with stable expected responses, which makes this a mismatch.
Question 17
A cloud engineering team must correct a data-processing weakness. Inventory answers may be reused for a few minutes, but stock levels change frequently throughout the day. Which implementation best satisfies the requirement? Choose ONE.
- cache key includes authorization context
- cache TTL aligned to source freshness
- semantic response cache
- precompute predictable responses
Correct Answer: B
Correct Answer
Answer B is correct because cache TTL aligned to source freshness is correct. It expire cached model responses when underlying business data may have changed. This matters because cached answers save cost but stale results would violate freshness requirements. That directly addresses the scenario.
Incorrect Answers
Answer A is incorrect because cache key includes authorization context helps when semantically identical questions can legitimately produce different answers for different users by allowing the pipeline to include tenant, permissions, or data-scope attributes in the key so cached content is not reused across security boundaries. The described issue is cached answers save cost but stale results would violate freshness requirements.
Answer C is incorrect because semantic response cache solves a different input problem. It helps to reuse a prior answer when a new request is semantically equivalent and freshness rules allow reuse, normally when users ask meaning-equivalent questions whose responses do not need a new FM invocation every time. The current workflow needs cached answers save cost but stale results would violate freshness requirements.
Answer D is incorrect because precompute predictable responses is appropriate where a set of common requests is predictable and can be prepared before users ask. It achieves that by helping to generate or retrieve stable high-demand answers ahead of request time. This scenario instead requires cached answers save cost but stale results would violate freshness requirements.
Question 18
A solutions architect is reviewing how source data reaches an FM. A chat UI waits several seconds for a long answer and currently displays nothing until the entire response arrives. What should be used to meet this requirement? Choose ONE.
- response streaming
- hybrid lexical and vector retrieval
- precompute predictable responses
- prompt caching for repeated long prefixes
Correct Answer: A
Correct Answer
Answer A is correct because Choose response streaming. Its role is to return generated tokens incrementally so the user sees progress before the full completion is finished. The key requirement is that perceived latency matters even when total generation time cannot be reduced enough.
Incorrect Answers
Answer B is incorrect because hybrid lexical and vector retrieval can help a pipeline combine keyword and semantic signals when exact identifiers and conceptual meaning are both important, especially when pure vector search misses exact product codes while pure keyword search misses paraphrased intent. The present requirement is perceived latency matters even when total generation time cannot be reduced enough, so it is not the best answer.
Answer C is incorrect because precompute predictable responses would improve another stage by enabling it to generate or retrieve stable high-demand answers ahead of request time. It fits when a set of common requests is predictable and can be prepared before users ask. The stated problem is perceived latency matters even when total generation time cannot be reduced enough.
Answer D is incorrect because prompt caching for repeated long prefixes works for cases where a substantial portion of input tokens is identical across many invocations by helping to reuse supported cached prompt context when many requests share a large stable instruction or document prefix. The present scenario needs perceived latency matters even when total generation time cannot be reduced enough.
Question 19
A platform team is reviewing a production GenAI data pipeline. A voice assistant must begin responding quickly and current model latency exceeds the user-experience target. What is the most appropriate implementation choice? Choose ONE.
- hybrid lexical and vector retrieval
- cache key includes authorization context
- latency-optimized model selection
- semantic response cache
Correct Answer: C
Correct Answer
Answer C is correct because latency-optimized model selection is appropriate because it will choose an approved model or inference option benchmarked for lower response latency when the use case is time sensitive. The requirement centers on a situation where interactive response time is more important than maximizing model sophistication.
Incorrect Answers
Answer A is incorrect because hybrid lexical and vector retrieval belongs in workflows where pure vector search misses exact product codes while pure keyword search misses paraphrased intent because it can combine keyword and semantic signals when exact identifiers and conceptual meaning are both important. This question instead asks for interactive response time is more important than maximizing model sophistication.
Answer B is incorrect because cache key includes authorization context is useful when semantically identical questions can legitimately produce different answers for different users and can include tenant, permissions, or data-scope attributes in the key so cached content is not reused across security boundaries. The described workload requires interactive response time is more important than maximizing model sophistication.
Answer D is incorrect because semantic response cache can reuse a prior answer when a new request is semantically equivalent and freshness rules allow reuse. It fits when users ask meaning-equivalent questions whose responses do not need a new FM invocation every time. This scenario instead requires interactive response time is more important than maximizing model sophistication, so it targets a different problem.
Question 20
An enterprise team is refining the input pipeline for a GenAI application. Users paste long conversational questions that include greetings and unrelated details before the actual technical issue. What should the developer implement? Choose ONE.
- query preprocessing before retrieval
- precompute predictable responses
- semantic response cache
- prompt caching for repeated long prefixes
Correct Answer: A
Correct Answer
Answer A is correct because The requirement points to query preprocessing before retrieval. It is meant to normalize or rewrite a noisy user query before searching the vector or hybrid index. The scenario needs an approach where retrieval quality suffers because raw queries contain irrelevant wording, misspellings, or weak search terms.
Incorrect Answers
Answer B is incorrect because precompute predictable responses helps when a set of common requests is predictable and can be prepared before users ask by allowing the pipeline to generate or retrieve stable high-demand answers ahead of request time. The described issue is retrieval quality suffers because raw queries contain irrelevant wording, misspellings, or weak search terms.
Answer C is incorrect because semantic response cache is relevant, but it is intended to reuse a prior answer when a new request is semantically equivalent and freshness rules allow reuse when users ask meaning-equivalent questions whose responses do not need a new FM invocation every time. The actual requirement is retrieval quality suffers because raw queries contain irrelevant wording, misspellings, or weak search terms.
Answer D is incorrect because prompt caching for repeated long prefixes solves a different input problem. It helps to reuse supported cached prompt context when many requests share a large stable instruction or document prefix, normally when a substantial portion of input tokens is identical across many invocations. The current workflow needs retrieval quality suffers because raw queries contain irrelevant wording, misspellings, or weak search terms.
