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.
Question 361
A retrieved paragraph begins 'This exception applies only to the above category,' but the category heading is absent. What should the pipeline do?
- Choose a category based on the most common customer question.
- Assume the exception applies to every category.
- Ignore the phrase about scope and quote only the exception.
- Retrieve the relevant heading or parent section before interpreting the exception.
Explanation
Correct Answer
D:The missing antecedent determines the exception's scope and is needed for a supported answer. A parent heading may be small but essential evidence. Retrieving it is more reliable than extending the exception to every category or guessing the missing antecedent.
Incorrect Answers
A:Popularity does not identify the category referenced by this document.
B:Universal application contradicts the paragraph's explicit scope restriction.
C:Omitting the restriction would misrepresent the source's meaning.
Question 362
The answer model cannot use the right policy because retrieval never returns it. Which metric most directly diagnoses this upstream problem?
- The number of tokens generated after retrieval.
- The proportion of retrieved passages that are relevant, without checking whether the required policy appears.
- Whether known relevant evidence appears among the retrieved candidates for representative queries.
- The proportion of final answers that include any citation, regardless of which source was retrieved.
Explanation
Correct Answer
C:A retrieval recall measure tests the stage responsible for supplying the necessary evidence. Use queries with known relevant sources to distinguish retrieval failure from generation failure. If the needed evidence never reaches the context, changing answer style will not fix the root cause.
Incorrect Answers
A:Output volume cannot show whether the correct source was retrieved.
B:That is a precision-oriented measure. A set can contain relevant passages yet omit the specific evidence needed; this failure requires checking retrieval recall for the needed policy.
D:Citation presence can coexist with failure to retrieve the necessary authoritative passage.
Question 363
A retrieval-augmented assistant answers incorrectly. Which THREE diagnostic checks help localize the failure? Choose THREE.
- Check whether the corpus contains the needed authoritative information.
- Inspect only the final answer's grammar and infer from fluent prose that retrieval worked.
- Rerun with a larger output limit and treat any longer answer as proof that the fault was fixed.
- Check whether generation faithfully used the evidence it received.
- Attribute the error to the model without inspecting which passages were available to it.
- Check whether retrieval selected the needed passage.
Explanation
Correct Answers
A:The retrieval system cannot return evidence that is absent from its accessible corpus.
D:Even with the correct passage present, synthesis may misinterpret or overstate it.
F:This distinguishes corpus availability from successful selection into the model's context.
Incorrect Answers
B:A well-written answer can still use missing or irrelevant evidence. This check does not distinguish corpus, retrieval, and generation failures.
C:Length is not evidence that the correct source entered the pipeline. The diagnostic must inspect the actual stage inputs and outputs.
E:The model may have received an incomplete or incorrect evidence set. Assigning cause before inspecting that boundary prevents reliable localization.
Question 364
A user loses access to a document that remains in a retrieval-result cache. What must the application do on subsequent requests?
- Hide the document title but still return the full content.
- Serve the cached text because it was authorized once.
- Apply current access controls before returning cached document content.
- Ask the model whether the user seems trustworthy.
Explanation
Correct Answer
C:Cache reuse must not bypass the user's current entitlement to the underlying source. The cache must preserve or recheck the access boundary rather than becoming an alternate disclosure path. Revocation can require invalidation or a current authorization check before reuse.
Incorrect Answers
A:Removing a title does not prevent disclosure of restricted text.
B:Past access does not establish current authorization after permission revocation.
D:A model's impression is not an access-control decision from the authoritative system.
Question 365
A policy document is withdrawn from the source repository but remains retrievable from the index. Which TWO lifecycle controls are needed? Choose TWO.
- Remove all document dates so age cannot affect retrieval.
- Propagate source deletion or withdrawal into the retrieval index.
- Treat indexed copies as permanently authoritative after first ingestion.
- Ask the answer model to guess which documents have been deleted.
- Verify that caches and derived retrieval stores stop serving the withdrawn material as current.
Explanation
Correct Answers
B:The index must reflect that the document is no longer an active source.
E:Downstream copies can preserve the stale source unless their lifecycle is reconciled too.
Incorrect Answers
A:Hiding time metadata makes stale content harder to identify.
C:Initial indexing does not override later withdrawal or correction by the source owner.
D:The model needs actual lifecycle metadata rather than an unsupported guess.
Question 366
A conversation summary says 'Offer refunds within 30 days' but drops the stated exception for final-sale items. What should the compaction review check?
- Whether the general 30-day rule remains present, without comparing its eligibility conditions.
- Only whether the summary is shorter than the original conversation.
- Whether the summary contains the words refunds and final-sale somewhere, regardless of their relationship.
- Whether qualifications that change permitted decisions survive the summary.
Explanation
Correct Answer
D:An omitted exception can alter future actions even when the general rule remains accurate. A compact summary can remain concise while keeping exceptions that affect eligibility. Correctness depends on retaining decision-changing details, not merely the most frequent statements.
Incorrect Answers
A:The general rule can be preserved while its exception is lost. Eligibility conditions are precisely what distinguishes allowed refunds from disallowed ones here.
B:Compression ratio does not reveal whether a decision-changing qualification was lost.
C:Keyword retention does not show that final-sale status still limits refund eligibility. The summary must preserve the operative relationship between the terms.
Question 367
An early message gives the delivery city as York; a later user correction says Leeds. What should the compacted task state record?
- Leeds as the current delivery city, with the correction relationship retained if needed.
- York only because it appeared first.
- Both cities as interchangeable options without noting the correction.
- Keep the city most frequently mentioned in the history, even when the later correction explicitly replaces it.
Explanation
Correct Answer
A:The summary should preserve the updated fact rather than revive superseded task state. The task state should distinguish superseded values from unresolved alternatives. Otherwise a later worker may mistakenly treat an already corrected address as still available for selection.
Incorrect Answers
B:The initial value was explicitly corrected and is no longer the current destination.
C:Unqualified alternatives reintroduce ambiguity that the user already resolved.
D:Frequency cannot override a clear correction. Repeated earlier references to York would not make it the current delivery destination.
Question 368
A custom agent waits until its history exceeds the model's context limit before requesting a summary of that entire history. The summary request is then rejected too. What should change?
- Trigger compaction before exhaustion, reserving enough capacity to process the history and produce the summary.
- Delete the oldest messages blindly, including unresolved constraints, whenever the limit is exceeded.
- Increase only the requested summary output length while keeping the oversized input.
- Retry the same oversized summary request without reducing its input.
Explanation
Correct Answer
A:The compaction operation itself needs a valid request budget. A proactive threshold lets the application preserve required task state instead of discovering the limit only after summarization can no longer run.
Incorrect Answers
B:Unselective truncation can remove the state needed for correct continuation rather than deliberately compacting it.
C:A larger output allowance does not remove the input-size violation and may leave even less usable budget.
D:Repeated submission does not make the history fit the model's input constraints.
Question 369
A question asks which retention policy applies to Project Harbor. The project registry identifies Harbor's owning division; a separate policy collection assigns policies by division and never mentions project names. A search for 'Harbor retention policy' finds no policy. What retrieval plan can answer the question from these sources?
- Retrieve Harbor's division from the registry, then retrieve that division's applicable policy and preserve both evidence links.
- Increase the number of project-name matches returned from the policy collection.
- Apply the highest-ranked policy returned by a broad search across all divisions.
- Treat the empty policy search as proof that Harbor has no retention requirements.
Explanation
Correct Answer
A:The answer depends on a join across two sources: project to division, then division to policy. The intermediate evidence makes the resulting applicability claim traceable.
Incorrect Answers
B:A larger result limit does not add the project names that this collection does not contain. The query needs the division identifier supplied by the other source.
C:Relevance to retention does not show that a policy governs Harbor. The missing division relationship must be established before selecting an applicable policy.
D:The policy collection is indexed by division, so the project's name need not occur in an applicable policy. The initial query did not resolve the relationship needed for retrieval.
Question 370
A retrieval service embeds queries with model B but compares them against document vectors created by incompatible model A. What should the index migration ensure?
- That the answer model is instructed to overlook irrelevant retrieved passages.
- Keep the old document vectors because they have the same number of dimensions as the new query vectors, without testing representation compatibility.
- That every query vector is sorted numerically before comparison.
- Query and document representations use a compatible embedding space, with a planned reindexing or parallel-index transition.
- Only that the vectors have similar-looking decimal values.
Explanation
Correct Answer
D:Nearest-neighbor comparisons require compatible representations rather than arbitrary vectors from different encoders. Even vectors with the same dimension can be incompatible if their encoders define different spaces. Reindexing or a controlled parallel transition should be evaluated before switching query embeddings.
Incorrect Answers
A:Generation instructions do not repair the upstream vector-space mismatch.
B:Equal dimensionality does not establish a shared semantic coordinate system. Compatible encoding is needed for meaningful similarity comparisons.
C:Sorting coordinates destroys the learned representation and cannot align incompatible spaces.
E:Visual similarity of numbers does not establish that their dimensions encode comparable meanings.