Question 240 A hybrid search retrieves a relevant policy at rank 15 while the generator sees only the top five chunks; increasing context size harms latency. Why does the current approach fail under these conditions? Coarse chunks can match topic vocabulary yet bury the actual answer span among unrelated context. A model's advertised context window constrains input and output budgeting; silently truncated evidence breaks grounding. Candidate recall may be adequate while final ranking fails; reranking addresses ordering without unbounded prompt expansion. Excessively fine segmentation removes procedural dependencies necessary for grounded explanations. ExplanationCorrect Answer C:Add selective reranking of a modest candidate pool and keep a fixed context budget rather than sending all fifteen passages. Other Answers A:Candidate recall may be adequate while final ranking fails; reranking addresses ordering without unbounded prompt expansion. B:Candidate recall may be adequate while final ranking fails; reranking addresses ordering without unbounded prompt expansion. D:That mechanism describes another problem: Tiny sentence chunks improve keyword precision but make an instruction about calibration useless because its required prerequisite sits one sentence earlier. It does not explain this case, where candidate recall may be adequate while final ranking fails; reranking addresses ordering without unbounded prompt expansion.
Question 241 A hybrid search retrieves a relevant policy at rank 15 while the generator sees only the top five chunks; increasing context size harms latency. Which observation would establish that the requirement is now met? Inspect retrieved calibration answers and verify prerequisites and measurement steps are present in a single sufficient evidence set. Compare answer support and P95 latency with and without reranking on the same labeled query set. Evaluate held-out warranty questions for support span accuracy and retrieval relevance across both chunking configurations. Count real prompt tokens on worst-case policies and verify final citations reference included—not truncated—text. ExplanationCorrect Answer B:Candidate recall may be adequate while final ranking fails; reranking addresses ordering without unbounded prompt expansion. Other Answers A:Compare answer support and P95 latency with and without reranking on the same labeled query set. C:Compare answer support and P95 latency with and without reranking on the same labeled query set. D:Compare answer support and P95 latency with and without reranking on the same labeled query set.
Question 242 A hybrid search retrieves a relevant policy at rank 15 while the generator sees only the top five chunks; increasing context size harms latency. What lasting control best prevents recurrence of the identified issue? Set maximum-context validation and retrieval fallback rules in the model-selection contract. Include cross-sentence dependency questions in the retrieval regression dataset. Establish separate ranking and latency acceptance gates for retrieval changes. Track source-level evidence quality, not only raw hit counts, when tuning retrieval. ExplanationCorrect Answer C:Candidate recall may be adequate while final ranking fails; reranking addresses ordering without unbounded prompt expansion. Other Answers A:Candidate recall may be adequate while final ranking fails; reranking addresses ordering without unbounded prompt expansion. B:This control addresses a different exposure: Tiny sentence chunks improve keyword precision but make an instruction about calibration useless because its required prerequisite sits one sentence earlier. It would not prevent this failure, because candidate recall may be adequate while final ranking fails; reranking addresses ordering without unbounded prompt expansion. D:That candidate recall may be adequate while final ranking fails; reranking addresses ordering without unbounded prompt expansion.
Question 243 A new embedding model changes similarity scores, and a fixed threshold rejects most valid questions even though the source corpus is unchanged. What is the most targeted corrective step for this observed failure? Increase chunk boundary awareness or bounded overlap so prerequisites and actions remain together for calibration queries. Rebalance metadata filters and lexical/vector hybrid search for exact code matches while measuring answer support. Recalibrate retrieval thresholds on labeled positives and negatives before promoting the new embedding model. Compare smaller section-aware chunks and a lower top-k using grounded-answer evaluations, selecting the configuration that improves citation precision. ExplanationCorrect Answer C:The action targets the stated cause: Similarity scales and distributions depend on the embedding model and cannot be assumed stable across versions. Check the repair by: plot positive/negative score distributions, adjust the threshold and verify recall plus false-positive rates. Other Answers A:Similarity scales and distributions depend on the embedding model and cannot be assumed stable across versions. B:Similarity scales and distributions depend on the embedding model and cannot be assumed stable across versions. D:Similarity scales and distributions depend on the embedding model and cannot be assumed stable across versions.
Question 244 A new embedding model changes similarity scores, and a fixed threshold rejects most valid questions even though the source corpus is unchanged. What is the decisive mechanism behind the observed result? Coarse chunks can match topic vocabulary yet bury the actual answer span among unrelated context. Pure semantic ranking can underweight rare exact identifiers when corpus density changes. Excessively fine segmentation removes procedural dependencies necessary for grounded explanations. Similarity scales and distributions depend on the embedding model and cannot be assumed stable across versions. ExplanationCorrect Answer D:Recalibrate retrieval thresholds on labeled positives and negatives before promoting the new embedding model. Other Answers A:Similarity scales and distributions depend on the embedding model and cannot be assumed stable across versions. B:That mechanism describes another problem: A document collection adds thousands of unrelated records, and previously good retrieval now shows poor grounding for rare product codes. It does not explain this case, where similarity scales and distributions depend on the embedding model and cannot be assumed stable across versions. C:Similarity scales and distributions depend on the embedding model and cannot be assumed stable across versions.
Question 245 A new embedding model changes similarity scores, and a fixed threshold rejects most valid questions even though the source corpus is unchanged. Which verification is strongest for this particular failure mode? Query held-out part numbers and compare exact-ID recall and groundedness before and after hybrid filtering. Inspect retrieved calibration answers and verify prerequisites and measurement steps are present in a single sufficient evidence set. Evaluate held-out warranty questions for support span accuracy and retrieval relevance across both chunking configurations. Plot positive/negative score distributions, adjust the threshold and verify recall plus false-positive rates. ExplanationCorrect Answer D:Similarity scales and distributions depend on the embedding model and cannot be assumed stable across versions. Other Answers A:Plot positive/negative score distributions, adjust the threshold and verify recall plus false-positive rates. B:Plot positive/negative score distributions, adjust the threshold and verify recall plus false-positive rates. C:Plot positive/negative score distributions, adjust the threshold and verify recall plus false-positive rates.
Question 246 A new embedding model changes similarity scores, and a fixed threshold rejects most valid questions even though the source corpus is unchanged. What change to engineering governance would reduce this risk over time? Treat embedding-model upgrades as re-indexing and evaluation events with threshold regression tests. Track source-level evidence quality, not only raw hit counts, when tuning retrieval. Monitor identifier coverage and apply corpus-drift checks when ingestion volume changes. Include cross-sentence dependency questions in the retrieval regression dataset. ExplanationCorrect Answer A:Similarity scales and distributions depend on the embedding model and cannot be assumed stable across versions. Other Answers B:This control addresses a different exposure: Long, broad chunks yield a high retrieval hit rate but the agent cites unrelated sections for narrow warranty questions. It would not prevent this failure, because similarity scales and distributions depend on the embedding model and cannot be assumed stable across versions. C:That similarity scales and distributions depend on the embedding model and cannot be assumed stable across versions. D:Similarity scales and distributions depend on the embedding model and cannot be assumed stable across versions.
Question 247 A document collection adds thousands of unrelated records, and previously good retrieval now shows poor grounding for rare product codes. Which engineering action most directly addresses this specific requirement? Classify the incident as a privacy/access-control failure, block unauthorized retrieval and validate output against disclosure rules. Increase chunk boundary awareness or bounded overlap so prerequisites and actions remain together for calibration queries. Compare smaller section-aware chunks and a lower top-k using grounded-answer evaluations, selecting the configuration that improves citation precision. Rebalance metadata filters and lexical/vector hybrid search for exact code matches while measuring answer support. ExplanationCorrect Answer D:The action targets the stated cause: Pure semantic ranking can underweight rare exact identifiers when corpus density changes. Check the repair by: query held-out part numbers and compare exact-ID recall and groundedness before and after hybrid filtering. Other Answers A:Pure semantic ranking can underweight rare exact identifiers when corpus density changes. B:Pure semantic ranking can underweight rare exact identifiers when corpus density changes. C:Pure semantic ranking can underweight rare exact identifiers when corpus density changes.
Question 248 A user writes “refund invoice INV-731,” but the assistant searches all invoices because it never extracts the invoice identifier. What is the most targeted corrective step for this observed failure? Resolve the entity against authorized account identifiers and ask for disambiguation rather than choosing a similar name. Fail the safety and tone criterion independently of policy factuality and route the case through controlled moderation or rewrite review. Extract date, time and location with timezone context, confirm ambiguous values, then invoke the availability tool. Parse invoice_id as a typed field, validate its format and restrict the authorized lookup to that exact invoice. ExplanationCorrect Answer D:The action targets the stated cause: Entity extraction provides a stable lookup key; searching by broad natural language can leak or misselect records. Check the repair by: test valid, malformed and missing IDs and inspect the resulting SQL parameters and permissions decisions. Other Answers A:Entity extraction provides a stable lookup key; searching by broad natural language can leak or misselect records. B:Entity extraction provides a stable lookup key; searching by broad natural language can leak or misselect records. C:Entity extraction provides a stable lookup key; searching by broad natural language can leak or misselect records.
Question 249 A user writes “refund invoice INV-731,” but the assistant searches all invoices because it never extracts the invoice identifier. What is the decisive mechanism behind the observed result? Correct factual content does not make abusive presentation acceptable; quality evaluation needs separate safety dimensions. Relative temporal language cannot be safely converted to a timestamp without locale and reference-date assumptions. Entity extraction provides a stable lookup key; searching by broad natural language can leak or misselect records. Free-form names are not stable unique keys and can refer to multiple records with different access controls. ExplanationCorrect Answer C:Parse invoice_id as a typed field, validate its format and restrict the authorized lookup to that exact invoice. Other Answers A:Entity extraction provides a stable lookup key; searching by broad natural language can leak or misselect records. B:Entity extraction provides a stable lookup key; searching by broad natural language can leak or misselect records. D:That mechanism describes another problem: A customer asks about the balance for “the Smith account,” and an agent inserts the first matching person into a finance prompt. It does not explain this case, where entity extraction provides a stable lookup key; searching by broad natural language can leak or misselect records.
Question 250 A user writes “refund invoice INV-731,” but the assistant searches all invoices because it never extracts the invoice identifier. Which verification is strongest for this particular failure mode? Score factuality and harmful-language behavior separately on adversarial tone prompts and confirm safe alternatives pass. Use two same-name fixture accounts and prove the agent abstains until an authorized unique account is selected. Test valid, malformed and missing IDs and inspect the resulting SQL parameters and permissions decisions. Test identical phrases in two timezones and verify no booking happens for ambiguous conversions. ExplanationCorrect Answer C:Entity extraction provides a stable lookup key; searching by broad natural language can leak or misselect records. Other Answers A:Test valid, malformed and missing IDs and inspect the resulting SQL parameters and permissions decisions. B:Test valid, malformed and missing IDs and inspect the resulting SQL parameters and permissions decisions. D:Test valid, malformed and missing IDs and inspect the resulting SQL parameters and permissions decisions.
