Generative AI Engineer Associate · 14% of the exam

Design applications: free practice questions

5 sample questions from our 35-question bank for this domain — answers and explanations included. These are the same scenario-based style as the real Databricks exam.

1. A fintech startup is designing an LLM-powered application to answer questions about SEC filings. The team wants to ensure the model never speculates about future stock prices and always cites the specific filing section it used. Which design elements should be included BEFORE the application is built? (Select TWO)

  • A. A system prompt guardrail that explicitly instructs the model not to generate forward-looking price predictions✓ Correct
  • B. Fine-tuning the model on historical SEC filings to improve domain vocabulary
  • C. A quality requirement specifying that every response must include a citation to the source filing section✓ Correct
  • D. A retrieval pipeline that fetches only the top-1 most similar chunk to minimize retrieved noise
  • E. A post-processing step that removes all numerical values from responses to prevent speculation
Explanation

The two correct answers are (A) and (C). (A) A system prompt guardrail is the standard mechanism for constraining prohibited behaviors — instructing the model upfront that speculative price predictions are off-limits. (C) Defining a citation quality requirement before building ensures the output format and grounding standard are baked into the prompt design and evaluation criteria from the start. (B) Fine-tuning on SEC filings may improve domain fluency but does not enforce behavioral guardrails or citation requirements. (D) Fetching only the top-1 chunk is a retrieval design choice and may hurt answer completeness; it is not a guardrail or quality requirement. (E) Removing all numbers post-hoc is too blunt and would destroy legitimate factual financial data in responses.

2. A company is deploying an LLM to help support agents draft responses to customer complaints. Stakeholders have identified three quality requirements: (1) responses must never promise refunds unless the agent adds a manual approval, (2) responses must stay under 150 words, and (3) the tone must be empathetic. Which design approach BEST addresses ALL THREE requirements simultaneously?

  • A. A system prompt that defines the empathetic tone persona, a word-count constraint, and an explicit prohibition on refund promises; combined with a programmatic post-processing step that flags any response containing refund language for human review✓ Correct
  • B. Fine-tune the model on approved past responses to implicitly encode tone, length, and refund policies
  • C. A system prompt for tone only, with separate manual review for refund promises and a separate summarization step to enforce word count
  • D. Use a lower temperature to make the model less likely to generate refund promises
Explanation

(A) is the best answer. Embedding all three constraints in the system prompt addresses tone, length, and the refund prohibition declaratively; the programmatic flagging step adds a hard safety layer for the highest-risk constraint (refund promises) that cannot be guaranteed by the LLM alone. This is a defense-in-depth design. (B) Fine-tuning can encode stylistic patterns but is not a reliable mechanism for enforcing strict business rules like the refund prohibition; it also requires labeled data and retraining cycles. (C) Using separate manual steps for each requirement is operationally fragmented, slow, and does not enforce requirements programmatically, leaving gaps in coverage. (D) Temperature controls randomness, not semantic content; a low-temperature model can still generate refund promises confidently.

3. A team is designing a multi-stage reasoning pipeline for a scientific research assistant. The pipeline must: synthesize findings from 20 retrieved papers, identify contradictions between studies, and produce a structured literature review with confidence scores. Which combination of chain components is MOST appropriate? (Select THREE)

  • A. A retrieval step that fetches the 20 most relevant paper abstracts and stores them as structured context✓ Correct
  • B. A map step that applies a per-paper extraction prompt to identify key claims and methodology from each abstract individually✓ Correct
  • C. A zero-shot single-call prompt that passes all 20 abstracts simultaneously and requests the full literature review
  • D. A reduce/synthesis step that consolidates the extracted claims, identifies contradictions, and drafts the structured review✓ Correct
  • E. A classification step that sorts papers into predefined topic buckets before any synthesis
  • F. A final scoring step that prompts the model to assign confidence scores to each synthesized finding based on the number of supporting studies
Explanation

The three correct components are (A), (B), and (D). (A) A structured retrieval step grounds the pipeline in real source material and formats it for downstream processing. (B) A map step processes each paper individually, extracting claims reliably without overwhelming the context window with all 20 papers at once. (D) A reduce/synthesis step consolidates the per-paper extractions, identifies contradictions across studies, and generates the structured review — the core reasoning step. (C) Passing all 20 abstracts in a single zero-shot call risks exceeding context limits, dilutes attention, and is less reliable than the map-reduce approach for contradiction detection. (E) A topic classification step is a useful organizational tool but is not required for the synthesis and contradiction-identification goals stated in the requirement. (F) A separate confidence scoring step is a plausible enhancement but the question asks for the MOST appropriate combination; confidence scoring is best handled as part of the reduce step (D) rather than as an additional stage, and selecting it over the three core architectural components would misrepresent the essential pipeline structure.

4. An e-learning platform is building a feature to automatically generate multiple-choice quiz questions from textbook chapters. The pedagogical team requires that: distractors must reflect common student misconceptions, correct answers must be traceable to a specific sentence in the chapter, and questions must be at a specified Bloom's taxonomy level. Which set of design decisions BEST meets these requirements? (Select TWO)

  • A. Use a few-shot prompt that includes two or three example questions with annotated distractors and Bloom's level labels to guide the model✓ Correct
  • B. Use zero-shot prompting and rely on the model's general knowledge of pedagogy
  • C. Include explicit output format instructions requiring the model to output the source sentence and Bloom's level tag alongside each question✓ Correct
  • D. Fine-tune a model on a public quiz dataset unrelated to the textbook domain
  • E. Use a chain-of-thought step that first identifies misconceptions and taxonomy levels before generating the question stem and options
  • F. Use a higher temperature setting to maximize output diversity without any structural constraints
Explanation

The two best design decisions are (A) and (C). (A) Few-shot examples that demonstrate annotated distractors and Bloom's level labels teach the model the exact output style required, making it far more reliable than zero-shot for a nuanced pedagogical format. (C) Explicit output format instructions requiring the source sentence and taxonomy tag ensure traceability and alignment with the pedagogical team's quality requirements. (B) Zero-shot prompting without examples is unlikely to consistently produce Bloom's-level-tagged questions with misconception-based distractors. (D) Fine-tuning on an unrelated public quiz dataset would introduce domain mismatch and would not embed the specific traceability or taxonomy requirements. (E) Chain-of-thought is potentially helpful but is not as directly tied to meeting the specific output structure requirements as (A) and (C). (F) High temperature without structural constraints would increase randomness and decrease the reliability of meeting the structured format requirements.

5. A prompt engineer is designing a system to answer complex multi-step math word problems. The model frequently gives wrong final answers when asked directly. Which technique is most likely to improve accuracy?

  • A. Zero-shot prompting with a longer system message
  • B. Few-shot prompting with chain-of-thought examples showing step-by-step reasoning✓ Correct
  • C. Increasing the model's temperature parameter
  • D. Using a retrieval-augmented generation pipeline to fetch similar problems
Explanation

Chain-of-thought (CoT) few-shot prompting, where examples demonstrate explicit intermediate reasoning steps, is well-documented to significantly improve accuracy on multi-step arithmetic and logical reasoning tasks. A longer system message without reasoning examples doesn't teach the model to decompose the problem. Increasing temperature introduces more randomness, which hurts deterministic math tasks. RAG helps when the model lacks factual knowledge, not when the problem is one of reasoning decomposition.

30 more questions in this domain

Practice the full bank with instant grading, flashcards, and a timed mock exam.

Start practicing free