Question 190 A re-ranker receives confidential chunks before caller entitlements are checked and places one unauthorized report first in the candidate list. Why does the current approach fail under these conditions? Cross-encoder scoring cost typically grows with query-candidate pairs; indiscriminate larger candidate lists increase processing time. Ranking is not a security control; scoring restricted passages still exposes them to a model or service before permission filtering. First-stage recall can include the right answer without placing it in the limited top-k; re-ranking changes ordering using richer query-passage signals. Corpus inclusion and access control determine what evidence a model can see; an answer prompt cannot reliably conceal already exposed confidential text. ExplanationCorrect Answer B:Filter candidates by user permissions before re-ranking, and independently enforce authorization at the underlying source service. Other Answers A:Ranking is not a security control; scoring restricted passages still exposes them to a model or service before permission filtering. C:That mechanism describes another problem: A hybrid search returns the correct maintenance manual at rank twelve, below generic articles that match keywords but not the failure code. It does not explain this case, where ranking is not a security control; scoring restricted passages still exposes them to a model or service before permission filtering. D:Ranking is not a security control; scoring restricted passages still exposes them to a model or service before permission filtering.
Question 191 A re-ranker receives confidential chunks before caller entitlements are checked and places one unauthorized report first in the candidate list. Which observation would establish that the requirement is now met? Compare relevant-document rank and grounded answer citations before and after the re-ranker on the same candidate pool. Plot recall and end-to-end latency as candidate count varies, then choose a validated budget under the service SLO. Probe with a low-privilege identity and show that restricted IDs never reach the re-ranker input or answer context. Inspect only final generated answers for restricted wording without examining which passages the re-ranker received. ExplanationCorrect Answer C:Ranking is not a security control; scoring restricted passages still exposes them to a model or service before permission filtering. Other Answers A:Probe with a low-privilege identity and show that restricted IDs never reach the re-ranker input or answer context. B:Probe with a low-privilege identity and show that restricted IDs never reach the re-ranker input or answer context. D:An answer may omit private text while unauthorized chunks were already sent into the ranking step. The re-ranker input must be checked using a low-privilege caller.
Question 192 A re-ranker receives confidential chunks before caller entitlements are checked and places one unauthorized report first in the candidate list. What lasting control best prevents recurrence of the identified issue? Require document classification, ACL propagation and access-filter regression tests for every source refresh. Benchmark two-stage retrieval with separate candidate-recall and final-ranking acceptance thresholds. Add latency and candidate-count guardrails to the retrieval pipeline with routine regression benchmarks. Treat access filtering as a pre-ranking invariant and test it with negative authorization fixtures. ExplanationCorrect Answer D:Ranking is not a security control; scoring restricted passages still exposes them to a model or service before permission filtering. Other Answers A:This control addresses a different exposure: A product warranty assistant indexes public brochures and confidential supplier negotiations together without separating who may read each. It would not prevent this failure, because ranking is not a security control; scoring restricted passages still exposes them to a model or service before permission filtering. B:That ranking is not a security control; scoring restricted passages still exposes them to a model or service before permission filtering. C:Ranking is not a security control; scoring restricted passages still exposes them to a model or service before permission filtering.
Question 193 A team increases the vector search candidate count from 40 to 1,000 while keeping a cross-encoder re-ranker, causing unacceptable response latency. What is the most targeted corrective step for this observed failure? Apply a relevance re-ranker to the initial candidate set using the full query and candidate passages before choosing generation context. Compare precision@k, recall@k, latency and token usage while tuning the cutoff; avoid treating higher recall as the sole goal. Filter candidates by user permissions before re-ranking, and independently enforce authorization at the underlying source service. Use a bounded candidate-recall stage and re-rank only a measured subset that preserves required recall at acceptable latency. ExplanationCorrect Answer D:The action targets the stated cause: Cross-encoder scoring cost typically grows with query-candidate pairs; indiscriminate larger candidate lists increase processing time. Check the repair by: plot recall and end-to-end latency as candidate count varies, then choose a validated budget under the service SLO. Other Answers A:Cross-encoder scoring cost typically grows with query-candidate pairs; indiscriminate larger candidate lists increase processing time. B:Cross-encoder scoring cost typically grows with query-candidate pairs; indiscriminate larger candidate lists increase processing time. C:Cross-encoder scoring cost typically grows with query-candidate pairs; indiscriminate larger candidate lists increase processing time.
Question 194 A team increases the vector search candidate count from 40 to 1,000 while keeping a cross-encoder re-ranker, causing unacceptable response latency. What is the decisive mechanism behind the observed result? First-stage recall can include the right answer without placing it in the limited top-k; re-ranking changes ordering using richer query-passage signals. Ranking is not a security control; scoring restricted passages still exposes them to a model or service before permission filtering. Cross-encoder scoring cost typically grows with query-candidate pairs; indiscriminate larger candidate lists increase processing time. Increasing k often captures additional relevant items at the cost of more irrelevant context and downstream generation expense. ExplanationCorrect Answer C:Use a bounded candidate-recall stage and re-rank only a measured subset that preserves required recall at acceptable latency. Other Answers A:That mechanism describes another problem: A hybrid search returns the correct maintenance manual at rank twelve, below generic articles that match keywords but not the failure code. It does not explain this case, where cross-encoder scoring cost typically grows with query-candidate pairs; indiscriminate larger candidate lists increase processing time. B:Cross-encoder scoring cost typically grows with query-candidate pairs; indiscriminate larger candidate lists increase processing time. D:Cross-encoder scoring cost typically grows with query-candidate pairs; indiscriminate larger candidate lists increase processing time.
Question 195 A team increases the vector search candidate count from 40 to 1,000 while keeping a cross-encoder re-ranker, causing unacceptable response latency. Which verification is strongest for this particular failure mode? Compare relevant-document rank and grounded answer citations before and after the re-ranker on the same candidate pool. Benchmark held-out questions at several k values and record both answer grounding and total token spend. Plot recall and end-to-end latency as candidate count varies, then choose a validated budget under the service SLO. Probe with a low-privilege identity and show that restricted IDs never reach the re-ranker input or answer context. ExplanationCorrect Answer C:Cross-encoder scoring cost typically grows with query-candidate pairs; indiscriminate larger candidate lists increase processing time. Other Answers A:Plot recall and end-to-end latency as candidate count varies, then choose a validated budget under the service SLO. B:Plot recall and end-to-end latency as candidate count varies, then choose a validated budget under the service SLO. D:Plot recall and end-to-end latency as candidate count varies, then choose a validated budget under the service SLO.
Question 196 A team increases the vector search candidate count from 40 to 1,000 while keeping a cross-encoder re-ranker, causing unacceptable response latency. What change to engineering governance would reduce this risk over time? Treat access filtering as a pre-ranking invariant and test it with negative authorization fixtures. Define a multi-metric retrieval acceptance budget and periodically retune top-k against representative traffic. Add latency and candidate-count guardrails to the retrieval pipeline with routine regression benchmarks. Benchmark two-stage retrieval with separate candidate-recall and final-ranking acceptance thresholds. ExplanationCorrect Answer C:Cross-encoder scoring cost typically grows with query-candidate pairs; indiscriminate larger candidate lists increase processing time. Other Answers A:That cross-encoder scoring cost typically grows with query-candidate pairs; indiscriminate larger candidate lists increase processing time. B:Cross-encoder scoring cost typically grows with query-candidate pairs; indiscriminate larger candidate lists increase processing time. D:This control addresses a different exposure: A hybrid search returns the correct maintenance manual at rank twelve, below generic articles that match keywords but not the failure code. It would not prevent this failure, because cross-encoder scoring cost typically grows with query-candidate pairs; indiscriminate larger candidate lists increase processing time.
Question 197 A team has three callable steps—retriever, prompt formatter and model—but its hand-written handler repeatedly loses passage IDs between steps. Which engineering action most directly addresses this specific requirement? Pin compatible framework and tool package versions, validate tool schemas and deploy from a reproducible environment with traceable builds. Select a framework with explicit conditional state transitions, tool interfaces and approval checkpoints instead of chaining every tool unconditionally. Choose a small deterministic job plus a model call; add an agent framework only when dynamic tool selection and stateful control are justified. Use an explicit chain composition abstraction with typed component inputs and outputs, passing citation metadata through every stage. ExplanationCorrect Answer D:The action targets the stated cause: Chain frameworks make intermediate transformations and dependencies inspectable; a plain concatenated string can discard structured provenance. Check the repair by: trace one request through each chain node and verify that answer citations map to the original retriever IDs. Other Answers A:Chain frameworks make intermediate transformations and dependencies inspectable; a plain concatenated string can discard structured provenance. B:Chain frameworks make intermediate transformations and dependencies inspectable; a plain concatenated string can discard structured provenance. C:Chain frameworks make intermediate transformations and dependencies inspectable; a plain concatenated string can discard structured provenance.
Question 198 A team has three callable steps—retriever, prompt formatter and model—but its hand-written handler repeatedly loses passage IDs between steps. What architectural constraint is most directly responsible? Library version drift can alter runnable interfaces or serialization even when model prompts are unchanged. A stateful agent graph can enforce branching and side-effect approvals where linear chains cannot safely represent the decision points. Chain frameworks make intermediate transformations and dependencies inspectable; a plain concatenated string can discard structured provenance. Framework complexity carries orchestration and debugging costs that a single fixed transformation does not require. ExplanationCorrect Answer C:Use an explicit chain composition abstraction with typed component inputs and outputs, passing citation metadata through every stage. Other Answers A:Chain frameworks make intermediate transformations and dependencies inspectable; a plain concatenated string can discard structured provenance. B:Chain frameworks make intermediate transformations and dependencies inspectable; a plain concatenated string can discard structured provenance. D:That mechanism describes another problem: A simple nightly text summarizer runs one API call, yet engineers propose a full autonomous agent framework with tools and memory. It does not explain this case, where chain frameworks make intermediate transformations and dependencies inspectable; a plain concatenated string can discard structured provenance.
Question 199 A team has three callable steps—retriever, prompt formatter and model—but its hand-written handler repeatedly loses passage IDs between steps. What validation evidence would most specifically confirm the resolution? Trace one request through each chain node and verify that answer citations map to the original retriever IDs. Run integration tests in the target runtime and compare tool-call payload schemas with the development environment. Replay read-only and write requests, confirming the correct tool route and zero writes without approval. Compare outputs, operational failure paths and execution cost of the minimal job against the agent prototype. ExplanationCorrect Answer A:Chain frameworks make intermediate transformations and dependencies inspectable; a plain concatenated string can discard structured provenance. Other Answers B:Trace one request through each chain node and verify that answer citations map to the original retriever IDs. C:Trace one request through each chain node and verify that answer citations map to the original retriever IDs. D:Trace one request through each chain node and verify that answer citations map to the original retriever IDs.
Question 200 A team has three callable steps—retriever, prompt formatter and model—but its hand-written handler repeatedly loses passage IDs between steps. What prevention measure is most targeted to this underlying cause? Document a complexity decision checklist that distinguishes fixed pipelines from actual agent tasks. Require trace coverage and sanitized observability as release criteria for agent implementations. Maintain automated component-contract tests and trace sampling whenever chain stages are reordered. Test graph invariants and authorization boundaries as part of every workflow release. ExplanationCorrect Answer C:Chain frameworks make intermediate transformations and dependencies inspectable; a plain concatenated string can discard structured provenance. Other Answers A:Chain frameworks make intermediate transformations and dependencies inspectable; a plain concatenated string can discard structured provenance. B:This control addresses a different exposure: A RAG application emits answers without a way to identify which tool or prompt stage produced an unsupported claim. It would not prevent this failure, because chain frameworks make intermediate transformations and dependencies inspectable; a plain concatenated string can discard structured provenance. D:That chain frameworks make intermediate transformations and dependencies inspectable; a plain concatenated string can discard structured provenance.
