Question 30 A company has a collection of expense memos and wants a one-sentence description of each memo that preserves the main expense purpose. What NLP task should drive the output? Generate a new memo about an unrelated expense. Return only whether the memo is written in English. Abstractive or extractive summarization constrained to one faithful sentence. Produce the embedding vector and display its raw numbers. ExplanationCorrect Answer C:The requirement is to condense each original memo while preserving intent rather than assign a category. Other Answers A:Creating different content is not a faithful one-sentence gist of the source. B:Language detection does not describe the expense purpose. D:Embedding coordinates are not a readable summary sentence.
Question 31 A marketplace receives item titles in two languages and needs standardized English titles while retaining model numbers. Which pairing of tasks best fits the need? Replace every identifier with a new randomly generated model code. Run sentiment analysis to determine whether each title is favorable. Use only semantic similarity rankings instead of producing English titles. Translate and normalize titles while explicitly preserving invariant product identifiers. ExplanationCorrect Answer D:Translation addresses language differences, and controlled normalization stabilizes wording without changing identifiers. Other Answers A:Invented codes break product identity and catalog joins. B:Sentiment is irrelevant to standardizing titles and identifiers. C:Similarity search does not create the requested normalized display values.
Question 32 A call-center manager wants a metric representing how satisfied callers feel in short feedback messages, rather than the subjects they discuss. Which text-analysis task is the closest match? Sentiment analysis with categories or a calibrated score, validated on domain-specific feedback. Entity extraction of every date and number as the final answer. RAG-based policy question answering for each comment. Unconditional translation of every already-English review. ExplanationCorrect Answer A:Sentiment labels assess expressed attitude; topic routing would instead classify subjects such as billing or delivery. Other Answers B:Dates and numbers do not directly measure satisfaction expressed in text. C:Answering policy questions is a different task from scoring expressed customer sentiment. D:Translation does not itself infer satisfaction.
Question 33 An HR team wants job applicants’ resumes assigned to one of several document types before separate, human-supervised screening. Which initial AI task matches this document-routing requirement? Produce a hiring verdict based only on sensitive applicant background traits. Document classification into the approved types with review of low-confidence or out-of-scope cases. Summarize every resume without producing the required document-type category. Generate images of resumes and choose labels from visual appearance alone. ExplanationCorrect Answer B:The immediate output is a discrete document class and not an autonomous hiring recommendation. Other Answers A:The task is document routing, not discriminatory or unsupervised candidate selection. C:A summary does not satisfy a routing label contract. D:Image generation cannot provide grounded category assignment from the content.
Question 34 A maintenance assistant must identify whether two short defect reports refer to the same component issue, though they use different synonyms. Which approach yields the most direct candidate evidence? Compare only the first three characters of each report. Use language detection as the sole matching criterion. Embed the report text and compare semantic similarity, then verify matches against component identifiers. Generate new defect text and ignore both original reports. ExplanationCorrect Answer C:Semantic vectors help find synonym-related reports; component-ID checks prevent semantically similar but different parts from being merged. Other Answers A:Prefix-character overlap is not a reliable test for component-level equivalence. B:Reports can share a language while describing different defects. D:Synthesizing unrelated reports discards the actual evidence needed for matching.
Question 35 A policy-chat application takes a user question and must answer only from the current approved handbook, citing passage IDs. Which ordered chain best matches this requirement? Generate an answer from memory first, then retrieve arbitrary passages to decorate it with citations. Retrieve every document across every tenant before checking user permissions. Skip source retrieval and use only a creative-writing prompt template. Authorize the user, retrieve relevant approved handbook passages, assemble cited context, then generate and validate the answer. ExplanationCorrect Answer D:Authorization and retrieval precede generation so the model sees relevant, permitted text, and the final response can be checked against citations. Other Answers A:Post-hoc citations may not support an already-invented answer. B:Global retrieval can expose unauthorized records before the access decision. C:No authoritative handbook context remains to ground the requested answer.
Question 36 A service needs to find policy passages for a question before calling a generator. A developer places the LLM completion step ahead of retrieval. What is the primary chain defect? The generator lacks retrieved policy evidence at answer time; retrieval should supply context before completion. Retrieval must always execute after all answers are published. Embedding the user query requires the generator to finish first. A postprocessor alone can add a missing policy rule without examining the document. ExplanationCorrect Answer A:A RAG answer’s factual grounding depends on relevant passages being available when the answer is constructed. Other Answers B:Post-publication retrieval cannot reliably ground a prior answer. C:Query embeddings can be produced directly from the question without pre-generating an answer. D:Formatting cannot reconstruct a current source rule that was never retrieved.
Question 37 A user asks an assistant for the balance of an internal purchase order. The system has an authenticated SQL tool and should avoid answering from a retrieved stale PDF. What chain is most reliable? Summarize an outdated PDF copy of the purchase order as the balance source. Extract and validate the order ID, authorize its row-level access, query the system of record, then format the returned balance. Ask the LLM to estimate balance from the order ID syntax. Call the SQL tool with an arbitrary table scan before checking access. ExplanationCorrect Answer B:A live keyed query produces authoritative current state, while authorization and validation prevent accidental cross-account disclosure. Other Answers A:The document may not reflect current settlements and is explicitly stale. C:Identifiers do not encode the current financial amount. D:Unscoped access can return confidential rows beyond the user’s entitlements.
Question 38 A chain must compute an invoice tax after extracting taxable amounts; the LLM sometimes miscalculates percentages. Which composition minimizes the calculation defect? Ask the model to guess the tax amount with more imaginative decoding. Generate a tax paragraph first and accept its embedded number without validation. Extract typed amounts, validate them, compute tax with deterministic code, and let the LLM describe the verified result. Translate the invoice before computing tax but never parse its numeric values. ExplanationCorrect Answer C:Separating arithmetic from language generation reduces hallucinated computations while preserving a readable final answer. Other Answers A:Higher randomness does not provide mathematically reliable arithmetic. B:The failure is numerical accuracy, which requires checking or deterministic calculation. D:Translation alone does not implement the required arithmetic calculation.
Question 39 A document assistant has a retriever that outputs passage objects and a generator that expects one context string with citation identifiers. What transform is required between the two steps? Convert every passage into a new model-training dataset before each query. Discard citation IDs and concatenate only the first words of each passage. Send raw database credentials as part of the context string. Format retrieved passage text and citation IDs into an escaped, size-bounded context payload for the generator. ExplanationCorrect Answer D:The transform adapts object-shaped retrieval results to the generator’s input contract and keeps provenance available for citations. Other Answers A:Per-request retraining is not the data-shape transform needed by the chain. B:That destroys provenance and likely omits relevant evidence. C:Credentials are unrelated sensitive data that should not be placed in a model prompt.
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.
