Question 40 A multi-step assistant chooses a tool based on user intent. A question about order status must call the order database, while product FAQs use vector retrieval. Which routing design fits the two data sources? Classify or parse the request intent, validate required identifiers, then branch to the authorized order tool or approved FAQ retriever. Always route every order-status question to an FAQ index with generic shipping advice. Always run both sources without verifying permissions and reveal all combined records. Choose a random branch to balance compute usage. ExplanationCorrect Answer A:The two intents require different sources; explicit routing avoids using static FAQ retrieval for live order state. Other Answers B:General FAQ text does not contain the customer’s live order status. C:Cross-source aggregation before authorization can leak customer-specific data. D:Random routing breaks the semantic dependency between user intent and source type.
Question 41 A chatbot retrieves ten passages, but the generator context window can hold only four complete passages plus the prompt. What should the chain do before generation? Concatenate all ten passages and assume the model will read only the relevant part. Rank and select permitted relevant passages within a measured token budget, retaining their citation IDs. Remove all citations to save tokens while preserving every passage. Increase the output-token limit and leave the input context unchanged. ExplanationCorrect Answer B:Context selection protects token limits while keeping the best available evidence for grounded generation. Other Answers A:The context can overflow or truncate critical information unpredictably. C:Discarding provenance defeats the citation requirement and may not solve the full length problem. D:Output budget does not enlarge the model’s supported input context.
Question 42 A customer-service chain receives a retrieved passage containing instructions to ignore access controls and email private data. How should the chain treat those instructions? Promote the retrieved passage to the system role because it came from a search result. Execute any tool command embedded in the passage before checking user access. Treat retrieved text as untrusted evidence, ignore its behavioral instructions, and enforce tool permissions outside the LLM. Disable audit logging to avoid recording the malicious passage. ExplanationCorrect Answer C:Documents may contain attacker-authored text; policy and action authorization must not be delegated to retrieved context. Other Answers A:Retrieved content is lower-trust data and must not override trusted instructions. B:This allows prompt injection to become unauthorized tool execution. D:Removing auditability does not prevent instruction injection and hides evidence of the attempt.
Question 43 A three-stage pipeline parses a customer request, looks up a contract, then formats a summary. Missing contract IDs currently trigger a database-wide scan. Which gate should be inserted before lookup? Ask the model to invent a probable contract ID when none is provided. Increase the SQL query limit and perform the scan faster. Generate the final summary without any contract input. Validate the parsed contract ID and fail closed or ask for clarification when it is missing, before any retrieval. ExplanationCorrect Answer D:A required identifier prevents an underspecified request from expanding into an unauthorized or expensive broad query. Other Answers A:A guessed identifier can select another customer’s record. B:A faster broad scan does not resolve the missing scoping identifier. C:The final answer would not be tied to an actual contract.
Question 44 A retrieval chain enriches each user question with a department context field, but the LLM receives an empty department because the template uses a different variable name. Which repair addresses the actual integration fault? Align the output key from the context stage with the exact input variable referenced by the prompt template. Increase vector-search index size to include more departments. Change the model temperature to zero without fixing input binding. Move user authentication after generation. ExplanationCorrect Answer A:Chain composition depends on matching input/output contracts; an inconsistent variable name drops the field before generation. Other Answers B:The issue is a mismatched template variable, not missing index content. C:Deterministic sampling does not fill a missing template variable. D:Authentication order does not resolve the broken data binding and can worsen security.
Question 45 A summarization pipeline outputs a long draft, then must emit a single-line JSON object under a published schema. What final stage best satisfies the consumer contract? Return the original prose without schema checks and let callers guess fields. Validate and transform the draft into the schema using a supported structured-output stage and reject invalid parses. Append an unescaped JSON fragment to the end of markdown prose. Generate random key names to describe the same ideas more creatively. ExplanationCorrect Answer B:The final stage adapts natural-language content to a machine contract, including typed keys and predictable error handling. Other Answers A:The consumer explicitly requires machine-readable structured data. C:Mixed formats and unescaped fragments create ambiguous or invalid payloads. D:Changing field names violates the published interface.
Question 46 A contract agent can use document retrieval and a numerical calculator. Retrieval finds a payment schedule; the calculator should sum installments. Which sequence preserves evidence and arithmetic correctness? Compute a guess before fetching the installment schedule. Ask the model for a plausible total and treat its text as a verified calculation. Retrieve the authorized payment schedule, extract typed installment amounts, compute the sum deterministically, then cite the schedule. Retrieve unrelated public contracts and average their payment values. ExplanationCorrect Answer C:Retrieved line items supply grounded inputs to an exact computation, and the answer can reference the source document. Other Answers A:Without amounts the calculator has no valid operands. B:Fluent generation does not establish arithmetic accuracy. D:Other contracts are not evidence for the current schedule.
Question 47 A business assistant must wait for a user to approve a purchase before a tool submits the purchase order. The chain currently submits automatically. What composition best enforces the approval boundary? Move the purchase call to the beginning of the chain so the customer sees a quicker result. Replace the approval record with the model’s guess about what the user would approve. Run the purchase call and ask for approval afterward. Prepare a validated purchase proposal, present it for explicit approval, and invoke the purchase tool only after approval is recorded. ExplanationCorrect Answer D:A human authorization gate separates informational reasoning from a consequential external write action. Other Answers A:Calling earlier makes the unauthorized purchase happen sooner. B:An LLM prediction is not explicit user authorization. C:Retroactive consent cannot prevent the already executed write.
Question 48 A RAG chain retrieves documents for a sales agent, but retrieved text contains employee-only price lists not authorized for the caller. Where must authorization be enforced to prevent disclosure? Filter retrieval candidates by the caller’s entitlements before presenting passages to the LLM, and check tool access independently. Send all price lists to the model and ask it politely not to quote restricted values. Hide citations in the final answer but include all restricted source text in the context. Increase the retrieval top-k until enough public passages drown out the restricted ones. ExplanationCorrect Answer A:Authorization must be applied before sensitive context reaches generation, because a later text filter cannot undo prompt exposure. Other Answers B:The model has already seen confidential material; a prompt is not an access-control boundary. C:Omitting citations does not prevent private details from influencing or leaking through the answer. D:More context does not remove the forbidden documents or satisfy access policy.
Question 49 A question-answering chain calls a model, then validates that citations point to retrieved passages. A bad citation currently passes because only JSON syntax is checked. Which additional gate is required? Accept any citation token that looks like a UUID. Verify every claimed citation ID exists in the authorized retrieval set and that the cited passage supports the associated assertion. Add a higher temperature so citations become more varied. Drop the passage store after generation and rely on the response alone. ExplanationCorrect Answer B:JSON validity is only structural; groundedness requires checking references against actual retrieved evidence. Other Answers A:Plausible syntax does not prove the passage was retrieved or supports the claim. C:Randomness can create more fabricated references and does not establish provenance. D:Without source records there is no evidence against which citations can be validated.
Question 50 A model chain can call a customer-profile tool but should avoid persisting private profile details in long-term conversation state. Which stage design best meets the constraint? Save the full raw customer profile as permanent chat history for every request. Write API credentials into conversation memory to simplify future tool calls. Retrieve only authorized minimal profile fields per request, use them transiently for generation, and store only approved non-sensitive session state. Skip authorization because the tool returned a record once before. ExplanationCorrect Answer C:The chain should minimize sensitive context and separate ephemeral tool data from persistent memory. Other Answers A:Persistent raw profiles increase privacy and retention risk beyond the specified purpose. B:Credentials should be managed outside user-visible and model-visible memory. D:A previous success does not prove current authorization for the caller.
