Question 350
A developer estimates token use from the user message alone, but requests still exceed the budget. Which TWO components should the preflight count also consider? Choose TWO.
- Only cached-token billing discounts, subtracting those tokens from logical context use.
- System instructions and tool definitions included in the request.
- Only the final answer's character count from an earlier unrelated task.
- Conversation history and tool results actually sent with the request.
- Only the filename of each input document instead of its included content.
Explanation
Correct Answers
B:These prompt components contribute tokens even though they are not part of the visible user message.
D:Prior messages and observations can occupy substantial context beyond the latest user text.
Incorrect Answers
A:A billing discount does not mean the cached prompt content ceases to occupy the model's request context.
C:An unrelated response length does not measure this request's input footprint.
E:A filename does not represent the tokens consumed by document text sent to the model.
Question 351
A service uses prompt caching and then omits a required policy document from a later request, expecting the cache to supply it. What is the mistake?
- Caching reuses eligible submitted prompt content; it does not replace constructing the required request context.
- A cache hit proves the omitted policy was included in the logical request.
- The model automatically remembers every document ever cached by the account.
- Caching always changes the policy's meaning before storage.
Explanation
Correct Answer
A:The application must still assemble the logical input the task needs rather than rely on hidden document recall. A cached prefix can be reused efficiently when the request is constructed appropriately. It does not grant a later unrelated request access to arbitrary documents previously processed by the account.
Incorrect Answers
B:A hit on another prefix does not establish that absent task content became available.
C:Prompt caching is not an account-wide conversational memory mechanism.
D:The problem is not an inherent semantic rewrite of cached content.
Question 352
A team enables prompt caching and wants to know whether repeated requests actually reuse cached input. What should it inspect?
- Only whether two responses use the same first word.
- Only the model's statement that it remembered the previous request.
- Only the application's count of repeated requests, without service usage data.
- Cache-related usage metrics together with latency and cost observations for comparable requests.
Explanation
Correct Answer
D:Reported usage distinguishes cache creation and reads, while operational measurements show their practical effect. Measure comparable repeated requests because changes in prompt composition can alter eligibility. Similar answers alone are not proof that the service performed a cache read.
Incorrect Answers
A:Output similarity does not demonstrate that input caching occurred.
B:Conversational wording is not evidence of a service-level cache hit.
C:Repeated input is an opportunity for reuse, not proof that cache eligibility and a hit occurred.
Question 353
An application sends the same eligible prompt after its cached entry has expired. What should the application expect?
- A normal processing path that may recreate the cache, with different cache-usage and latency characteristics.
- The model will necessarily lose the ability to read that text.
- Return the previous generated answer directly whenever the prompt text matches, assuming input prompt caching is the same feature as response caching.
- Keep reporting a cache hit from the previous call without inspecting the current response's usage metrics.
- The request must fail because the old cache no longer exists.
Explanation
Correct Answer
A:Expiration changes performance reuse, not the requirement that the valid request remain processable. An expired entry can turn a cache read into ordinary processing or a new cache creation. The source content is still present in the submitted request, so correctness should not depend on the cache surviving.
Incorrect Answers
B:The request still provides the content; expiration affects reuse rather than basic input comprehension.
C:Prompt reuse does not imply replay of a prior output. The request remains an inference operation unless a separate response-cache mechanism is implemented.
D:A past hit does not establish reuse after expiration. The current request's metrics determine how this processing attempt used the cache.
E:Cache expiration should not make otherwise valid supplied input unavailable for normal processing.
Question 354
A retrieval pipeline cuts every policy document at a fixed character count, often separating a rule from its exception. Which change most directly addresses the problem?
- Make every chunk one character long.
- Increase answer verbosity without changing retrieved evidence.
- Remove all exceptions before indexing to simplify retrieval.
- Use meaningful sections with the context needed for qualifications.
Explanation
Correct Answer
D:Semantically coherent chunks reduce the chance that retrieval presents a rule without its relevant exception. Headings, exceptions, and definitions can determine how a paragraph applies. Chunk design should preserve those relationships without indiscriminately making every chunk the whole document.
Incorrect Answers
A:Extremely small fragments destroy the context needed to interpret policy meaning.
B:A longer answer cannot recover an exception absent from the selected context.
C:Deleting qualifications changes the policy and makes resulting answers unreliable.
Question 355
Overlapping retrieval chunks repeat the same transaction paragraph. A summarizer counts the transaction twice. Which TWO changes address this failure? Choose TWO.
- Count every returned chunk as an independent real-world event.
- Deduplicate the underlying transaction evidence before aggregation.
- Assume repeated wording always indicates a second transaction.
- Track document offsets or source spans so overlapping evidence can be recognized.
- Remove every overlapping chunk regardless of whether it contains unique evidence.
Explanation
Correct Answers
B:Aggregation should count source events, not repeated presentations of the same event.
D:Source-position metadata exposes when two chunks repeat the same underlying passage.
Incorrect Answers
A:Retrieval units are not necessarily distinct events and can overlap deliberately.
C:Identical text can be duplicated by chunking rather than by the source describing another event.
E:Blanket removal can discard useful boundary context and distinct content.
Question 356
A search returns five mirrors of one announcement and one independent contradictory report. What should the answer treat as evidence strength?
- Five independent confirmations of the announcement.
- One underlying announcement plus the independent conflicting report, with provenance retained.
- Resolve the conflict by counting all matching pages equally, without checking their common origin.
- The longest mirrored page as definitive proof.
Explanation
Correct Answer
B:Source lineage prevents duplicated distribution from being mistaken for independent corroboration. Copies can help locate an announcement but do not constitute independent verification of its claim. The contradictory report still needs to be assessed on its own evidence and authority.
Incorrect Answers
A:Mirrors of the same origin do not provide five independent observations.
C:Multiple mirrors are dependent copies of one announcement. Equal page votes would overstate that source's support relative to the independent report.
D:Length does not resolve independence or the substantive contradiction.
Question 357
Raising retrieval top-k from 10 to 100 reduces answer quality by flooding the prompt with weakly related passages. What should the team optimize?
- Relevance and coverage within a context budget, using selection or reranking rather than maximizing count.
- The largest possible top-k regardless of the query.
- The number of repeated passages to reinforce each point.
- Answer length alone while continuing to send all 100 weakly selected passages.
Explanation
Correct Answer
A:Useful retrieval balances necessary evidence against distraction and available context. The evaluation should test whether necessary passages survive selection, not just whether token use decreases. A smaller set is useful only if it retains enough evidence to answer the question.
Incorrect Answers
B:More passages are not automatically more useful and can crowd out relevant material.
C:Redundant evidence consumes context without adding coverage.
D:Shortening output does not repair distracting or incomplete selection of the evidence supplied to the model.
Question 358
A first-stage search retrieves 50 plausible passages, but only a few directly answer the question. What is the role of a reranker?
- To reorder candidates by their relevance to the query before selecting context.
- To manufacture new source facts for passages with low relevance.
- Use reranking to recover an authoritative passage that was never included in the first-stage candidate set, without performing any additional retrieval.
- To guarantee every corpus document appears in the prompt.
- To replace source access controls with similarity scores.
Explanation
Correct Answer
A:A second-stage relevance assessment can focus the limited prompt on stronger evidence. Reranking cannot recover a document that the first stage never retrieved. Candidate recall and final selection quality therefore need separate checks.
Incorrect Answers
B:Ranking should reorder evidence, not invent content that the corpus does not contain.
C:A reranker operates on the candidates it receives. Missing evidence requires better initial retrieval or another search stage.
D:A reranker selects or orders candidates and need not include the complete corpus.
E:Relevance ranking does not determine whether the user is authorized to see a document.
Question 359
A retrieval query rewriter changes 'Which plans do not include phone support?' to 'Plans with phone support.' What went wrong?
- The original query should always be discarded after rewriting.
- The retrieved answers can be trusted if their titles contain support.
- The rewrite lost the negation and changed the information need.
- The query became shorter, which always improves retrieval.
Explanation
Correct Answer
C:A query expansion must preserve the original intent, including exclusion conditions. Negation, date bounds, and entity constraints are part of the retrieval intent. A rewritten query should broaden useful matching without reversing those conditions.
Incorrect Answers
A:Retaining the original can help detect or recover from meaning-changing rewrites.
B:Keyword presence does not establish relevance to the user's negative condition.
D:Shortness does not compensate for reversing the user's requested condition.
Question 360
Semantic search often misses an exact error code such as ZX-481 while retrieving conceptually related errors. What retrieval design should be evaluated?
- Remove the error code from every query to make it more abstract.
- Combine lexical matching for exact identifiers with semantic retrieval for related descriptions.
- Replace all error codes in the corpus with the same placeholder.
- Increase the semantic similarity threshold without adding an exact-token retrieval path.
Explanation
Correct Answer
B:Complementary retrieval methods can preserve exact-token matches while covering paraphrased explanations. Identifiers can be poorly served by semantic similarity even when their surrounding descriptions are meaningful. Combining retrieval signals should be evaluated against cases requiring both exact codes and paraphrased concepts.
Incorrect Answers
A:Dropping the discriminating identifier can make exact-match retrieval even less likely.
C:Erasing identifiers prevents the system from distinguishing the relevant error.
D:A stricter threshold may discard the relevant identifier match rather than solve semantic retrieval's blind spot.