Databricks GenAI Engineer Application Development: 292 practice questions

12 of the 292 Application Development questions in the Certsqill Databricks GenAI Engineer bank, shown in full below. Each one carries an explanation for every option, not just the correct one — the wrong answers are where the marks go.

Preparing for Databricks GenAI Engineer? Take the free 5-min readiness check →

1. Track the artifact: Which implementation meets the requirement?

1. Track the artifact, dependencies, signature, example, and trace. ✓

   This records every evidence type required by the exhibit for reproducibility and serving validation.
2. Save notebook output and describe inputs in a project note.

   An informal note does not provide the required tracked dependencies, signature, input example, and run trace.
3. Log only the final artifact.

   An artifact alone omits dependencies, interface expectations, and execution evidence.
4. Save the artifact and dependency file, then rerun the notebook without tracking the signature, input example, or execution trace.

   This omits three required evidence types even though it preserves the artifact and dependencies.

Track the artifact, dependencies, signature, input example, and run trace.

2. Map `query` to `question`: What change resolves this specific defect?

1. Add more retrieval documents first.

   More documents cannot fix a template variable mismatch that fails before model invocation.
2. Replace the model with a larger one.

   Model size is unrelated to the unresolved `question` variable.
3. Move context into the system prompt and remove retrieval.

   This changes the application and discards retrieval context without fixing the missing `question` variable.
4. Map `query` to `question`. ✓

   The template requires `question`, while the caller provides `query`; mapping the names fixes the pre-invocation failure.

Map `query` to the template’s `question` variable.

3. Offline references assess quality: Which interpretation is correct?

1. Inference records automatically score qualitative correctness for every response.

   Records capture requests and serving information; correctness scoring requires suitable scorers, labels, or review.
2. Reference answers are unnecessary because observed outputs can independently establish correctness, latency, errors, drift, and production task success.

   Observed outputs support operational analysis but cannot independently establish correctness without evaluation expectations or suitable review.
3. Offline references assess quality; online traces diagnose live behavior. ✓

   References support controlled quality evaluation, while traces support operational investigation of deployed behavior.
4. Production traces replace reference answers after enough requests.

   Production volume does not automatically provide expectations for judging correctness.

Offline references assess quality; online traces diagnose production behavior.

4. Resolve aliases to canonical entities before enrichment: What prerequisite should the team add before entity-a

1. Add a reranker that scores the complete source table after retrieval enrichment.

   This does not provide the required canonical mapping or perform the missing entity-aware query enrichment step.
2. Resolve aliases to canonical entities before enrichment. ✓

   A canonical alias-resolution layer maps abbreviations to the entities represented in the index before query expansion occurs.
3. Add a larger generator context window to every request.

   Generation capacity does not identify supplier aliases or improve entity matching before retrieval enrichment.
4. Change the search button label.

   A user-interface label does not map abbreviated mentions to canonical indexed entities.

Resolve supplier aliases to canonical entities before enriching the query.

5. Model B, because it satisfies both measured thresholds: Which choice meets both thresholds?

1. Model A, because its latency is comfortably below the threshold.

   Model A meets latency but fails the required groundedness threshold of 0.85.
2. Choose the cheapest model and increase retrieval depth until groundedness reaches 0.85.

   The supplied measurements already show Model A below the quality threshold, and retrieval changes require new evaluation.
3. Model C, because it has the highest groundedness score.

   Model C meets groundedness but exceeds the p95 latency limit by 150 milliseconds.
4. Model B, because it satisfies both measured thresholds. ✓

   Model B has 780 ms p95 latency and 0.89 groundedness, satisfying both requirements.

Model B is the only measured alternative within both the p95 latency and groundedness thresholds.

6. Choose a commercially permitted model: Which change resolves the licensing defect?

1. Add more evaluation examples to the prompt.

   Prompt examples can affect behavior but cannot expand the model license to permit commercial use.
2. Choose a commercially permitted model. ✓

   Selecting a model licensed for the intended commercial use resolves the licensing defect, subject to its obligations.
3. Lower the model temperature and document the deployment as production-ready.

   Runtime settings and deployment documentation do not change research-only licensing restrictions.
4. Create a production alias.

   An alias changes version management, not the model’s license or permitted uses.

Choose a model licensed for the intended commercial use.

7. Rechunk below 512 tokens: Which architecture change is correct?

1. Rechunk below 512 tokens, then regenerate embeddings and update the index. ✓

   Chunks must fit the embedding window, and changed chunk contents require new vectors and affected index updates.
2. Keep chunks unchanged and rely on the reranker to repair truncated embeddings.

   A reranker cannot reliably restore information lost when oversized text is truncated before embedding.
3. Switch to any embedding model with the same vector dimension.

   Equal dimensions do not make embedding spaces interchangeable; the index requires compatible encoders and vectors.
4. Increase generator output tokens.

   Generator output capacity does not change the embedding model’s 512-token input limit.

Rechunk within the embedding window, regenerate compatible vectors, and update the index.

8. Compare each answer claim directly with the retrieved: Before changing the model or prompt, what should the te

1. Fine-tune the assistant on additional catalog conversations containing return questions.

   Fine-tuning may change behavior, but it does not establish whether this response is supported by policy evidence.
2. Compare each answer claim directly with the retrieved policy and score its groundedness. ✓

   The unsupported thirty-day claim should be checked against retrieved evidence before selecting a remediation.
3. Add a stricter JSON schema requiring every response to contain a citation field.

   A citation field can enforce format, but it cannot determine whether the answer is supported by retrieved evidence.
4. Increase retrieval depth so the assistant can generate a more detailed returns explanation.

   More context may add detail, but the supplied policy already contradicts the unsupported answer claim.

Check the claim against retrieved policy evidence and measure groundedness before changing system components.

9. Compare chunking variants on held-out recall: Which implementation best meets both constraints?

1. Compare chunking variants on held-out recall, p95 latency, and token use. ✓

   This ablation tests rare-identifier recall while measuring both hard operational constraints before selection.
2. Add a reranker while retaining current chunks, and measure only final answer accuracy.

   A reranker cannot recover identifiers absent from the candidate set, and answer accuracy does not test latency or context use.
3. Use hybrid retrieval with more candidates and accept extra prompt context when recall improves.

   The added prompt context directly violates the fixed context-budget requirement, even if hybrid retrieval improves recall.
4. Increase chunk size and maximize recall.

   Optimizing recall alone can exceed both the p95 latency limit and the fixed context budget.

Ablate chunking strategies against recall, p95 latency, and token use.

10. The new prompt is overrestricting paraphrased evidence: The edit added, “Answer only when the exact wording ap

1. The new prompt is overrestricting paraphrased evidence. ✓

   The exact-wording instruction explains increased refusals despite unchanged grounded-answer quality.
2. The output schema is rejecting valid answers after generation.

   No schema change or parsing failure is reported; the refusal instruction itself explains the behavior.
3. The retriever must be returning fewer documents.

   The evidence reports a prompt change and stable grounded quality, not reduced retrieval volume or recall.
4. The model requires fine-tuning before it can follow retrieved instructions.

   The model already produced grounded answers, so the observed refusal increase does not establish a training deficiency.

The exact-match instruction rejects valid paraphrases, creating false refusals without improving grounding.

11. Model A fails the release gate on task success despite: Select TWO valid conclusions.

Select two. More than one option is correct — every correct one is ticked below.

1. Model A is release-eligible because its lower latency and cost offset its lower task-success score.

   Neither lower latency nor lower cost overrides the explicit minimum task-success requirement, which Model A fails.
2. Model B’s higher task-success score establishes that it will cost less in production.

   The measured costs show the opposite, and task success does not establish an unmeasured cost advantage.
3. Model A fails the release gate on task success despite meeting the latency limit. ✓

   Its 1.8-second p95 latency passes the latency requirement, but its 0.88 task-success score is below the required 0.90.
4. Both models pass because each satisfies at least one of the two thresholds.

   The release requirements are conjunctive: a model must satisfy both thresholds, not merely one.
5. Model B is the only model that satisfies both release thresholds. ✓

   Model B has 0.91 task success and 2.4-second p95 latency, while Model A has only 0.88 success and fails the success threshold.

Model B alone passes both thresholds; Model A fails on task success.

12. Require confirmation by an authorized manager before: What prerequisite remains before enabling the action?

1. Put the manager-approval requirement only in the prompt instructions.

   Prompt text alone is not an enforceable approval control for a consequential action.
2. Require confirmation by an authorized manager before submission. ✓

   Managerial confirmation supplies the required approval for this consequential external action and should be enforced before submission.
3. Proceed after the automated checks finish.

   The completed checks do not replace the policy-required approval from an authorized manager.
4. Add more retrieved examples of previously approved discounts.

   Examples may provide context but cannot authorize the current customer discount request.

Require confirmation from an authorized manager before submission.

Other Databricks GenAI Engineer domains

• Assembling and Deploying Applications — 212 questions →
• Design Applications — 136 questions →
• Data Preparation — 133 questions →
• Evaluation and Monitoring — 116 questions →
• Governance — 78 questions →
• All 967 Databricks GenAI Engineer questions →
• Databricks GenAI Engineer certification: requirements, cost and exam format →