Topic 12 Practice Test 1 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
A cloud engineering team must correct a data-processing weakness. A chat service sometimes assembles very large prompts and needs to estimate token consumption before calling the model. What is the most appropriate implementation choice? Choose ONE.
- response schema with concise fields
- context pruning by relevance
- prompt compression
- token counting before invocation
Correct Answer: D
Correct Answer
Answer D is correct because token counting before invocation is the direct remedy. It can estimate prompt and expected completion tokens before sending the request so oversized or unexpectedly expensive calls can be controlled. The critical scenario condition is that the application needs predictable cost and context-window use before FM inference.
Incorrect Answers
Answer A is incorrect because response schema with concise fields belongs in workflows where free-form generation includes explanations the downstream system discards because it can request a bounded structured output containing only fields the application consumes. This question instead asks for the application needs predictable cost and context-window use before FM inference.
Answer B is incorrect because context pruning by relevance is useful when prompt context has grown over time and includes material unrelated to the current question and can remove low-value history and retrieved passages while preserving information needed for the current request. The described workload requires the application needs predictable cost and context-window use before FM inference.
Answer C is incorrect because prompt compression targets a different issue by helping to rewrite verbose instructions and repeated context into a shorter representation that preserves essential meaning. It fits where prompt tokens are dominated by redundant wording rather than necessary evidence. The deciding condition is the application needs predictable cost and context-window use before FM inference.
Question 2
A solutions architect is reviewing how source data reaches an FM. A status assistant should return no more than a short paragraph, but the model sometimes generates several pages. What should the developer implement? Choose ONE.
- maximum response-token limit
- token-usage telemetry by feature
- conversation summarization for long history
- prompt compression
Correct Answer: A
Correct Answer
Answer A is correct because Use maximum response-token limit. It will cap generation length so routine responses cannot consume an excessive number of completion tokens. The deciding condition is that the business requires concise answers and wants a hard upper bound on completion cost. This matches the requirement.
Incorrect Answers
Answer B is incorrect because token-usage telemetry by feature helps when cost reduction requires knowing which application paths consume the most tokens by allowing the pipeline to record prompt and completion usage per workflow so the team can identify the largest cost drivers. The described issue is the business requires concise answers and wants a hard upper bound on completion cost.
Answer C is incorrect because conversation summarization for long history is relevant, but it is intended to replace older detailed turns with a compact maintained summary while retaining recent conversation verbatim when long-running chats need continuity without resending every historical token. The actual requirement is the business requires concise answers and wants a hard upper bound on completion cost.
Answer D is incorrect because prompt compression solves a different input problem. It helps to rewrite verbose instructions and repeated context into a shorter representation that preserves essential meaning, normally when prompt tokens are dominated by redundant wording rather than necessary evidence. The current workflow needs the business requires concise answers and wants a hard upper bound on completion cost.
Question 3
A platform team is reviewing a production GenAI data pipeline. A knowledge assistant retrieves 20 chunks, but offline evaluation shows the top five consistently contain the required evidence. Which implementation provides the required behavior? Choose ONE.
- retrieval top-k reduction after quality testing
- response schema with concise fields
- context pruning by relevance
- conversation summarization for long history
Correct Answer: A
Correct Answer
Answer A is correct because retrieval top-k reduction after quality testing directly fits. It can send fewer high-value retrieved chunks when evaluation shows additional chunks do not improve answer quality. The scenario is specifically about the case where RAG context is expensive because too many marginal passages are inserted into every prompt.
Incorrect Answers
Answer B is incorrect because response schema with concise fields has value because it can request a bounded structured output containing only fields the application consumes. Its proper fit is when free-form generation includes explanations the downstream system discards. This question requires RAG context is expensive because too many marginal passages are inserted into every prompt.
Answer C is incorrect because context pruning by relevance can help a pipeline remove low-value history and retrieved passages while preserving information needed for the current request, especially when prompt context has grown over time and includes material unrelated to the current question. The present requirement is RAG context is expensive because too many marginal passages are inserted into every prompt, so it is not the best answer.
Answer D is incorrect because conversation summarization for long history would improve another stage by enabling it to replace older detailed turns with a compact maintained summary while retaining recent conversation verbatim. It fits when long-running chats need continuity without resending every historical token. The stated problem is RAG context is expensive because too many marginal passages are inserted into every prompt.
Question 4
An enterprise team is refining the input pipeline for a GenAI application. A system prompt contains multiple paragraphs that restate the same output and safety requirements. Which change should the team make? Choose ONE.
- token-usage telemetry by feature
- structured concise system prompt
- response schema with concise fields
- prompt compression
Correct Answer: B
Correct Answer
Answer B is correct because The correct choice is structured concise system prompt. It helps to remove duplicated prose and express stable constraints in a compact structured instruction block. The deciding detail is that system instructions are unnecessarily verbose and repeated on each request.
Incorrect Answers
Answer A is incorrect because token-usage telemetry by feature can record prompt and completion usage per workflow so the team can identify the largest cost drivers, which matters when cost reduction requires knowing which application paths consume the most tokens. The present case is driven by system instructions are unnecessarily verbose and repeated on each request.
Answer C is incorrect because response schema with concise fields is useful when free-form generation includes explanations the downstream system discards and can request a bounded structured output containing only fields the application consumes. The described workload requires system instructions are unnecessarily verbose and repeated on each request.
Answer D is incorrect because prompt compression targets a different issue by helping to rewrite verbose instructions and repeated context into a shorter representation that preserves essential meaning. It fits where prompt tokens are dominated by redundant wording rather than necessary evidence. The deciding condition is system instructions are unnecessarily verbose and repeated on each request.
Question 5
An application team is standardizing data before foundation-model inference. A product assistant adds a 20-page policy context to greetings and simple navigation questions. Which approach most directly addresses the stated constraint? Choose ONE.
- context pruning by relevance
- conditional context inclusion
- conversation summarization for long history
- token-usage telemetry by feature
Correct Answer: B
Correct Answer
Answer B is correct because Select conditional context inclusion. Its purpose is to attach expensive reference context only for requests that actually need it. The operational requirement is that all requests currently carry a large reference bundle even when simple questions can be answered without it.
Incorrect Answers
Answer A is incorrect because context pruning by relevance is designed to remove low-value history and retrieved passages while preserving information needed for the current request. That makes sense when prompt context has grown over time and includes material unrelated to the current question. The question is centered on all requests currently carry a large reference bundle even when simple questions can be answered without it.
Answer C is incorrect because conversation summarization for long history helps when long-running chats need continuity without resending every historical token by allowing the pipeline to replace older detailed turns with a compact maintained summary while retaining recent conversation verbatim. The described issue is all requests currently carry a large reference bundle even when simple questions can be answered without it.
Answer D is incorrect because token-usage telemetry by feature is relevant, but it is intended to record prompt and completion usage per workflow so the team can identify the largest cost drivers when cost reduction requires knowing which application paths consume the most tokens. The actual requirement is all requests currently carry a large reference bundle even when simple questions can be answered without it.
Question 6
An operations team is improving the reliability of an FM input path. A classification endpoint uses the highest-cost model for simple category assignment with stable accuracy requirements. Which option best meets the operational requirement? Choose ONE.
- fallback escalation to a larger model
- complexity-based tiered model routing
- small-model routing for simple requests
- cost-aware route with quality floor
Correct Answer: C
Correct Answer
Answer C is correct because The strongest answer is small-model routing for simple requests. It works by helping to send low-complexity tasks to a less expensive capable model while reserving larger models for harder work. The requirement is that many requests do not require the capability or price of the most powerful FM.
Incorrect Answers
Answer A is incorrect because fallback escalation to a larger model addresses another concern by helping to start with an efficient model and escalate only low-confidence or failed cases to a more capable model. It fits where most requests are easy but a minority need stronger reasoning. Here the deciding need is many requests do not require the capability or price of the most powerful FM.
Answer B is incorrect because Using complexity-based tiered model routing would let the team classify requests by difficulty and select a model tier that balances quality and inference price. That is appropriate when workload complexity varies significantly and a single model tier wastes cost on easy cases. The scenario instead calls for many requests do not require the capability or price of the most powerful FM.
Answer D is incorrect because cost-aware route with quality floor can help a pipeline choose the cheapest eligible model only after filtering candidates that meet the required quality threshold, especially when the optimization objective is cost but quality cannot fall below a contractual minimum. The present requirement is many requests do not require the capability or price of the most powerful FM, so it is not the best answer.
Question 7
A developer is hardening an FM-consumption workflow. Two models have different prices and accuracy. The team needs the lowest-cost option that still meets its acceptance score. Which solution most directly resolves this issue? Choose ONE.
- on-demand inference for sporadic traffic
- quality-adjusted price benchmarking
- separate real-time and batch model paths
- fallback escalation to a larger model
Correct Answer: B
Correct Answer
Answer B is correct because quality-adjusted price benchmarking is correct. It compare candidate models using measured task quality together with per-request cost rather than price alone. This matters because the cheapest model may require retries or fail quality thresholds, so selection must consider value. That directly addresses the scenario.
Incorrect Answers
Answer A is incorrect because on-demand inference for sporadic traffic would be sound if the goal were request volume is low and bursty with long idle periods; it can use pay-per-request capacity when traffic is intermittent and does not justify continuously reserved throughput. The requirement here is the cheapest model may require retries or fail quality thresholds, so selection must consider value.
Answer C is incorrect because separate real-time and batch model paths belongs in workflows where one architecture currently forces both user-facing and nightly processing through the same synchronous path because it can route interactive and offline workloads to different inference patterns optimized for their latency and cost needs. This question instead asks for the cheapest model may require retries or fail quality thresholds, so selection must consider value.
Answer D is incorrect because fallback escalation to a larger model is useful when most requests are easy but a minority need stronger reasoning and can start with an efficient model and escalate only low-confidence or failed cases to a more capable model. The described workload requires the cheapest model may require retries or fail quality thresholds, so selection must consider value.
Question 8
A GenAI workload is moving from prototype to production. A platform runs extraction, summarization, and complex planning workloads through the same expensive model. Which design is the best fit for this requirement? Choose ONE.
- cost-aware route with quality floor
- on-demand inference for sporadic traffic
- task-specific model mapping
- complexity-based tiered model routing
Correct Answer: C
Correct Answer
Answer C is correct because Choose task-specific model mapping. Its role is to assign different approved models to summarization, extraction, and reasoning based on benchmarked capability. The key requirement is that one model is not the best price-performance choice for every task type.
Incorrect Answers
Answer A is incorrect because cost-aware route with quality floor targets a neighboring concern. It can choose the cheapest eligible model only after filtering candidates that meet the required quality threshold when the optimization objective is cost but quality cannot fall below a contractual minimum. The key requirement here is one model is not the best price-performance choice for every task type.
Answer B is incorrect because on-demand inference for sporadic traffic is designed to use pay-per-request capacity when traffic is intermittent and does not justify continuously reserved throughput. That makes sense when request volume is low and bursty with long idle periods. The question is centered on one model is not the best price-performance choice for every task type.
Answer D is incorrect because complexity-based tiered model routing helps when workload complexity varies significantly and a single model tier wastes cost on easy cases by allowing the pipeline to classify requests by difficulty and select a model tier that balances quality and inference price. The described issue is one model is not the best price-performance choice for every task type.
Question 9
A regulated workflow requires deterministic preprocessing before inference. A customer-facing assistant receives a stable high request rate around the clock with predictable token volume. Which implementation best satisfies the requirement? Choose ONE.
- provisioned throughput for steady high utilization
- fallback escalation to a larger model
- cost-aware route with quality floor
- separate real-time and batch model paths
Correct Answer: A
Correct Answer
Answer A is correct because provisioned throughput for steady high utilization is appropriate because it will reserve model throughput when sustained predictable demand can justify committed capacity and reduce capacity uncertainty. The requirement centers on a situation where traffic is consistently high and utilization can keep reserved capacity productive.
Incorrect Answers
Answer B is incorrect because fallback escalation to a larger model addresses another concern by helping to start with an efficient model and escalate only low-confidence or failed cases to a more capable model. It fits where most requests are easy but a minority need stronger reasoning. Here the deciding need is traffic is consistently high and utilization can keep reserved capacity productive.
Answer C is incorrect because Using cost-aware route with quality floor would let the team choose the cheapest eligible model only after filtering candidates that meet the required quality threshold. That is appropriate when the optimization objective is cost but quality cannot fall below a contractual minimum. The scenario instead calls for traffic is consistently high and utilization can keep reserved capacity productive.
Answer D is incorrect because separate real-time and batch model paths has value because it can route interactive and offline workloads to different inference patterns optimized for their latency and cost needs. Its proper fit is when one architecture currently forces both user-facing and nightly processing through the same synchronous path. This question requires traffic is consistently high and utilization can keep reserved capacity productive.
Question 10
During a production-readiness review, engineers identify a data-handling issue. A nightly catalog enrichment job processes 200,000 descriptions and users never wait for individual results. What should be used to meet this requirement? Choose ONE.
- separate real-time and batch model paths
- complexity-based tiered model routing
- on-demand inference for sporadic traffic
- batch inference for offline workloads
Correct Answer: D
Correct Answer
Answer D is correct because The requirement points to batch inference for offline workloads. It is meant to process noninteractive requests in batches rather than paying the latency overhead of individual synchronous calls. The scenario needs an approach where large jobs have no real-time response requirement and can be grouped efficiently.
Incorrect Answers
Answer A is incorrect because separate real-time and batch model paths is appropriate where one architecture currently forces both user-facing and nightly processing through the same synchronous path. It achieves that by helping to route interactive and offline workloads to different inference patterns optimized for their latency and cost needs. This scenario instead requires large jobs have no real-time response requirement and can be grouped efficiently.
Answer B is incorrect because complexity-based tiered model routing would be sound if the goal were workload complexity varies significantly and a single model tier wastes cost on easy cases; it can classify requests by difficulty and select a model tier that balances quality and inference price. The requirement here is large jobs have no real-time response requirement and can be grouped efficiently.
Answer C is incorrect because on-demand inference for sporadic traffic can use pay-per-request capacity when traffic is intermittent and does not justify continuously reserved throughput, which matters when request volume is low and bursty with long idle periods. The present case is driven by large jobs have no real-time response requirement and can be grouped efficiently.
Question 11
A cloud engineering team must correct a data-processing weakness. An offline evaluation job sends tens of thousands of tiny independent prompts one request at a time. What is the most appropriate implementation choice? Choose ONE.
- adaptive worker concurrency
- autoscaling from queue depth and processing rate
- token-throughput capacity planning
- request batching for compatible work
Correct Answer: D
Correct Answer
Answer D is correct because request batching for compatible work matches the constraint. It can combine compatible small inference items into supported batches to improve throughput and reduce per-request overhead. This is the intended approach when many independent small items can be processed together without user-facing latency requirements.
Incorrect Answers
Answer A is incorrect because adaptive worker concurrency works for cases where optimal parallelism changes as token sizes and service conditions change by helping to change the number of simultaneous inference workers using observed latency and throttling rather than a permanently high setting. The present scenario needs many independent small items can be processed together without user-facing latency requirements.
Answer B is incorrect because autoscaling from queue depth and processing rate targets a neighboring concern. It can scale consumers based on backlog and measured token-processing capacity instead of fixed instance counts when asynchronous GenAI work accumulates in a queue and demand changes throughout the day. The key requirement here is many independent small items can be processed together without user-facing latency requirements.
Answer C is incorrect because token-throughput capacity planning is designed to size inference capacity using observed input and output token rates rather than request count alone. That makes sense when requests vary widely in prompt and completion size, making raw requests-per-second misleading. The question is centered on many independent small items can be processed together without user-facing latency requirements.
Question 12
A solutions architect is reviewing how source data reaches an FM. A worker pool launches hundreds of Bedrock calls at once and spends significant time retrying throttled requests. What should the developer implement? Choose ONE.
- concurrency control at the caller
- off-peak scheduling for flexible batch jobs
- autoscaling from queue depth and processing rate
- parallelization of independent subrequests
Correct Answer: A
Correct Answer
Answer A is correct because Use concurrency control at the caller for this design. It will limit simultaneous model invocations to a level the downstream quota and application can sustain. That aligns with the need that unbounded parallelism creates throttling and retries that reduce effective throughput.
Incorrect Answers
Answer B is incorrect because off-peak scheduling for flexible batch jobs is useful when the workload has a completion deadline but does not need to run during the busiest interactive window because it can move nonurgent high-volume inference to periods with lower competing demand when business timing permits. The current requirement is unbounded parallelism creates throttling and retries that reduce effective throughput, which makes this a mismatch.
Answer C is incorrect because autoscaling from queue depth and processing rate addresses another concern by helping to scale consumers based on backlog and measured token-processing capacity instead of fixed instance counts. It fits where asynchronous GenAI work accumulates in a queue and demand changes throughout the day. Here the deciding need is unbounded parallelism creates throttling and retries that reduce effective throughput.
Answer D is incorrect because Using parallelization of independent subrequests would let the team run independent model or retrieval operations concurrently when their results do not depend on one another. That is appropriate when serial execution adds latency even though several workflow branches have no dependency. The scenario instead calls for unbounded parallelism creates throttling and retries that reduce effective throughput.
Question 13
A platform team is reviewing a production GenAI data pipeline. A provisioned model deployment is either underused for weeks or regularly reaches its throughput limit. Which implementation provides the required behavior? Choose ONE.
- token-throughput capacity planning
- parallelization of independent subrequests
- provisioned-throughput utilization monitoring
- adaptive worker concurrency
Correct Answer: C
Correct Answer
Answer C is correct because provisioned-throughput utilization monitoring is the best fit because it can track reserved-capacity usage and adjust commitments when utilization remains too low or saturation remains too high. The workflow needs a case where the team needs to avoid both idle reserved capacity and chronic capacity pressure.
Incorrect Answers
Answer A is incorrect because token-throughput capacity planning solves a different input problem. It helps to size inference capacity using observed input and output token rates rather than request count alone, normally when requests vary widely in prompt and completion size, making raw requests-per-second misleading. The current workflow needs the team needs to avoid both idle reserved capacity and chronic capacity pressure.
Answer B is incorrect because parallelization of independent subrequests is appropriate where serial execution adds latency even though several workflow branches have no dependency. It achieves that by helping to run independent model or retrieval operations concurrently when their results do not depend on one another. This scenario instead requires the team needs to avoid both idle reserved capacity and chronic capacity pressure.
Answer D is incorrect because adaptive worker concurrency can change the number of simultaneous inference workers using observed latency and throttling rather than a permanently high setting, which matters when optimal parallelism changes as token sizes and service conditions change. The present case is driven by the team needs to avoid both idle reserved capacity and chronic capacity pressure.
Question 14
An enterprise team is refining the input pipeline for a GenAI application. A document-upload service receives a large burst after business hours and downstream inference quotas are fixed. Which change should the team make? Choose ONE.
- backpressure with durable queueing
- autoscaling from queue depth and processing rate
- adaptive worker concurrency
- off-peak scheduling for flexible batch jobs
Correct Answer: A
Correct Answer
Answer A is correct because Implement backpressure with durable queueing. It is designed to accept work into a queue and process it at sustainable throughput instead of letting producers overwhelm inference capacity. Here, the important constraint is that incoming demand can temporarily exceed available model throughput but requests must not be lost.
Incorrect Answers
Answer B is incorrect because autoscaling from queue depth and processing rate works for cases where asynchronous GenAI work accumulates in a queue and demand changes throughout the day by helping to scale consumers based on backlog and measured token-processing capacity instead of fixed instance counts. The present scenario needs incoming demand can temporarily exceed available model throughput but requests must not be lost.
Answer C is incorrect because adaptive worker concurrency targets a neighboring concern. It can change the number of simultaneous inference workers using observed latency and throttling rather than a permanently high setting when optimal parallelism changes as token sizes and service conditions change. The key requirement here is incoming demand can temporarily exceed available model throughput but requests must not be lost.
Answer D is incorrect because off-peak scheduling for flexible batch jobs is designed to move nonurgent high-volume inference to periods with lower competing demand when business timing permits. That makes sense when the workload has a completion deadline but does not need to run during the busiest interactive window. The question is centered on incoming demand can temporarily exceed available model throughput but requests must not be lost.
Question 15
An application team is standardizing data before foundation-model inference. A nightly summarization job causes latency spikes for the customer chat service because both share the same capacity path. Which approach most directly addresses the stated constraint? Choose ONE.
- off-peak scheduling for flexible batch jobs
- token-throughput capacity planning
- parallelization of independent subrequests
- separate capacity pools by workload priority
Correct Answer: D
Correct Answer
Answer D is correct because separate capacity pools by workload priority should be used. It can isolate high-priority interactive traffic from background jobs so batch demand cannot consume all available throughput. That capability is needed because user-facing requests require predictable performance while offline jobs are flexible.
Incorrect Answers
Answer A is incorrect because off-peak scheduling for flexible batch jobs targets a different issue by helping to move nonurgent high-volume inference to periods with lower competing demand when business timing permits. It fits where the workload has a completion deadline but does not need to run during the busiest interactive window. The deciding condition is user-facing requests require predictable performance while offline jobs are flexible.
Answer B is incorrect because token-throughput capacity planning can size inference capacity using observed input and output token rates rather than request count alone. It fits when requests vary widely in prompt and completion size, making raw requests-per-second misleading. This scenario instead requires user-facing requests require predictable performance while offline jobs are flexible, so it targets a different problem.
Answer C is incorrect because parallelization of independent subrequests is useful when serial execution adds latency even though several workflow branches have no dependency because it can run independent model or retrieval operations concurrently when their results do not depend on one another. The current requirement is user-facing requests require predictable performance while offline jobs are flexible, which makes this a mismatch.
Question 16
An operations team is improving the reliability of an FM input path. Employees repeatedly ask slightly different phrasings of the same stable benefits-policy question. Which option best meets the operational requirement? Choose ONE.
- cache TTL aligned to source freshness
- semantic response cache
- deterministic request-hash cache
- latency-optimized model selection
Correct Answer: B
Correct Answer
Answer B is correct because semantic response cache solves the right problem. It will reuse a prior answer when a new request is semantically equivalent and freshness rules allow reuse. The case depends on the fact that users ask meaning-equivalent questions whose responses do not need a new FM invocation every time.
Incorrect Answers
Answer A is incorrect because cache TTL aligned to source freshness is relevant, but it is intended to expire cached model responses when underlying business data may have changed when cached answers save cost but stale results would violate freshness requirements. The actual requirement is users ask meaning-equivalent questions whose responses do not need a new FM invocation every time.
Answer C is incorrect because deterministic request-hash cache is appropriate where the workload receives byte-equivalent or canonically equivalent requests with stable expected responses. It achieves that by helping to cache results by a canonical hash when identical deterministic inputs should always reuse the same output. This scenario instead requires users ask meaning-equivalent questions whose responses do not need a new FM invocation every time.
Answer D is incorrect because latency-optimized model selection would be sound if the goal were interactive response time is more important than maximizing model sophistication; it can choose an approved model or inference option benchmarked for lower response latency when the use case is time sensitive. The requirement here is users ask meaning-equivalent questions whose responses do not need a new FM invocation every time.
Question 17
A developer is hardening an FM-consumption workflow. Every request includes the same large policy manual followed by a small user-specific question. Which solution most directly resolves this issue? Choose ONE.
- cache TTL aligned to source freshness
- response streaming
- prompt caching for repeated long prefixes
- query preprocessing before retrieval
Correct Answer: C
Correct Answer
Answer C is correct because prompt caching for repeated long prefixes is the direct remedy. It can reuse supported cached prompt context when many requests share a large stable instruction or document prefix. The critical scenario condition is that a substantial portion of input tokens is identical across many invocations.
Incorrect Answers
Answer A is incorrect because cache TTL aligned to source freshness can help a pipeline expire cached model responses when underlying business data may have changed, especially when cached answers save cost but stale results would violate freshness requirements. The present requirement is a substantial portion of input tokens is identical across many invocations, so it is not the best answer.
Answer B is incorrect because response streaming would improve another stage by enabling it to return generated tokens incrementally so the user sees progress before the full completion is finished. It fits when perceived latency matters even when total generation time cannot be reduced enough. The stated problem is a substantial portion of input tokens is identical across many invocations.
Answer D is incorrect because query preprocessing before retrieval targets a neighboring concern. It can normalize or rewrite a noisy user query before searching the vector or hybrid index when retrieval quality suffers because raw queries contain irrelevant wording, misspellings, or weak search terms. The key requirement here is a substantial portion of input tokens is identical across many invocations.
Question 18
A GenAI workload is moving from prototype to production. Two departments ask the same policy question but are allowed to see different confidential appendices. Which design is the best fit for this requirement? Choose ONE.
- deterministic request-hash cache
- latency-optimized model selection
- cache key includes authorization context
- response streaming
Correct Answer: C
Correct Answer
Answer C is correct because Use cache key includes authorization context. It will include tenant, permissions, or data-scope attributes in the key so cached content is not reused across security boundaries. The deciding condition is that semantically identical questions can legitimately produce different answers for different users. This matches the requirement.
Incorrect Answers
Answer A is incorrect because deterministic request-hash cache is useful when the workload receives byte-equivalent or canonically equivalent requests with stable expected responses and can cache results by a canonical hash when identical deterministic inputs should always reuse the same output. The described workload requires semantically identical questions can legitimately produce different answers for different users.
Answer B is incorrect because latency-optimized model selection targets a different issue by helping to choose an approved model or inference option benchmarked for lower response latency when the use case is time sensitive. It fits where interactive response time is more important than maximizing model sophistication. The deciding condition is semantically identical questions can legitimately produce different answers for different users.
Answer D is incorrect because response streaming is useful when perceived latency matters even when total generation time cannot be reduced enough because it can return generated tokens incrementally so the user sees progress before the full completion is finished. The current requirement is semantically identical questions can legitimately produce different answers for different users, which makes this a mismatch.
Question 19
A regulated workflow requires deterministic preprocessing before inference. Every morning thousands of employees request the same daily briefing generated from data that is finalized at 5 a.m. Which implementation best satisfies the requirement? Choose ONE.
- cache TTL aligned to source freshness
- query preprocessing before retrieval
- latency-optimized model selection
- precompute predictable responses
Correct Answer: D
Correct Answer
Answer D is correct because precompute predictable responses directly fits. It can generate or retrieve stable high-demand answers ahead of request time. The scenario is specifically about the case where a set of common requests is predictable and can be prepared before users ask.
Incorrect Answers
Answer A is incorrect because cache TTL aligned to source freshness helps when cached answers save cost but stale results would violate freshness requirements by allowing the pipeline to expire cached model responses when underlying business data may have changed. The described issue is a set of common requests is predictable and can be prepared before users ask.
Answer B is incorrect because query preprocessing before retrieval is relevant, but it is intended to normalize or rewrite a noisy user query before searching the vector or hybrid index when retrieval quality suffers because raw queries contain irrelevant wording, misspellings, or weak search terms. The actual requirement is a set of common requests is predictable and can be prepared before users ask.
Answer C is incorrect because latency-optimized model selection solves a different input problem. It helps to choose an approved model or inference option benchmarked for lower response latency when the use case is time sensitive, normally when interactive response time is more important than maximizing model sophistication. The current workflow needs a set of common requests is predictable and can be prepared before users ask.
Question 20
During a production-readiness review, engineers identify a data-handling issue. A support corpus contains error codes such as E417 alongside natural-language troubleshooting descriptions. What should be used to meet this requirement? Choose ONE.
- response streaming
- hybrid lexical and vector retrieval
- deterministic request-hash cache
- query preprocessing before retrieval
Correct Answer: B
Correct Answer
Answer B is correct because The correct choice is hybrid lexical and vector retrieval. It helps to combine keyword and semantic signals when exact identifiers and conceptual meaning are both important. The deciding detail is that pure vector search misses exact product codes while pure keyword search misses paraphrased intent.
Incorrect Answers
Answer A is incorrect because response streaming has value because it can return generated tokens incrementally so the user sees progress before the full completion is finished. Its proper fit is when perceived latency matters even when total generation time cannot be reduced enough. This question requires pure vector search misses exact product codes while pure keyword search misses paraphrased intent.
Answer C is incorrect because deterministic request-hash cache would improve another stage by enabling it to cache results by a canonical hash when identical deterministic inputs should always reuse the same output. It fits when the workload receives byte-equivalent or canonically equivalent requests with stable expected responses. The stated problem is pure vector search misses exact product codes while pure keyword search misses paraphrased intent.
Answer D is incorrect because query preprocessing before retrieval works for cases where retrieval quality suffers because raw queries contain irrelevant wording, misspellings, or weak search terms by helping to normalize or rewrite a noisy user query before searching the vector or hybrid index. The present scenario needs pure vector search misses exact product codes while pure keyword search misses paraphrased intent.
