Databricks GenAI Engineer Design Applications: 136 practice questions

12 of the 136 Design Applications 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. Authenticate first: Which implementation best meets the requirements?

1. Retrieve approved documents, generate an answer without using them as context, then authenticate and attach citations after generation.

   Authentication must precede retrieval, and the model must use retrieved context during synthesis; attaching citations afterward does not satisfy those requirements.
2. Authenticate, retrieve, and answer from the documents.

   This omits prompt assembly, structured-output parsing, and validation, so malformed responses may be returned.
3. Authenticate first, retrieve permitted documents, assemble the prompt, generate, parse, validate, and return citations. ✓

   This sequence checks identity before protected retrieval, grounds generation in approved documents, and validates the structured response before returning citations.
4. Authenticate, retrieve permitted documents, assemble the prompt, generate, return citations, and validate the response afterward.

   Validation occurs after the response is returned, so malformed structured output can already reach the consumer.

Authenticate before retrieval, use retrieved context for generation, and validate the structured response before returning citations.

2. The response is valid JSON but violates the declared field: What best explains the observed failure?

1. The model failed to retrieve the operations policy before generating JSON.

   The evidence identifies a type mismatch, not missing retrieval or unsupported policy context.
2. The parser incorrectly rejects every response containing a boolean field.

   The evidence shows rejection of a string value against a boolean schema, not all booleans.
3. The schema proves that the model’s operational decision is factually incorrect, so generation must be retrained before validation.

   Schema validation detects structural or type violations; it cannot establish whether the operational decision is true.
4. The response is valid JSON but violates the declared field types. ✓

   JSON syntax is valid, yet `retryable` contains a string where the contract requires a boolean.

The JSON parses syntactically, but `retryable` violates the boolean type required by the output schema.

3. Use Knowledge Assistant for citation-grounded: Which TWO choices best meet these requirements?

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

1. Use Knowledge Assistant for citation-grounded conversational answers over the contract collection. ✓

   Knowledge Assistant supports conversational question answering over documents with supporting citations.
2. Use Information Extraction to populate contract fields, then return those fields as the assistant’s cited answers.

   Information Extraction is intended for structured fields, not conversational answers with supporting passages for arbitrary questions.
3. Use specialist feedback to evaluate answer quality and guide subsequent improvements. ✓

   Specialist feedback provides human evaluation evidence for improving document-answer quality.
4. Fine-tune the model on the contracts so newly uploaded documents become its current factual memory.

   Fine-tuning changes learned behavior; it is not a dependable live store for newly uploaded contract facts.
5. Use Supervisor Agent as the sole document-answering component, without configuring document retrieval or citations.

   Supervisor Agent coordinates agents and tools but does not replace document retrieval and citation configuration.

Knowledge Assistant supplies citation-grounded document QA, and specialist feedback supports evaluation and improvement.

4. Quality criteria for accepting an answer: Before implementation begins, which missing contract element should

1. Quality criteria for accepting an answer. ✓

   The application still needs measurable quality criteria to determine whether generated responses meet requirements.
2. A dynamic agent that can select tools without bounded calls or validation.

   Unbounded tool selection conflicts with safe implementation and does not fill the missing quality requirement.
3. A second identity provider for every assistant request.

   The scenario already covers permissions; adding another identity provider is not a required contract element.
4. A fine-tuning dataset containing all future analytics questions and answers.

   Future questions are not required for defining the application contract, and fine-tuning is not automatically necessary.

Inputs, outputs, permissions, tools, and constraints still need measurable answer-quality acceptance criteria.

5. Choose per-user on-behalf-of authentication despite: Which alternative should be selected?

1. Choose per-user authentication without retrieval because security matters more than runbook context.

   This preserves identity but removes the retrieved knowledge required for incident triage answers.
2. Choose per-user on-behalf-of authentication despite its measured latency. ✓

   On-behalf-of authentication preserves user-specific access filtering, satisfying the security requirement despite slower retrieval.
3. Choose anonymous credentials for the lowest latency.

   Anonymous access is fast but fails the requirement to enforce each user’s runbook permissions.
4. Choose shared service credentials because they provide complete runbook coverage.

   Shared credentials can expose documents beyond the requesting user’s authorization boundary.

Per-user on-behalf-of authentication is the only measured choice combining correct access filtering with runbook retrieval.

6. Replace classification with retrieval-grounded answer: Which change directly resolves the defect?

1. Increase the response token limit while retaining the same classification output.

   A larger limit does not change classification into retrieval-grounded explanation generation.
2. Add more category labels to the existing classifier.

   Additional labels still produce classifications rather than the requested catalog answer and supporting source text.
3. Fine-tune the classifier until it emits longer category names.

   Longer labels remain classifications and do not reliably supply current catalog facts or source support.
4. Replace classification with retrieval-grounded answer generation. ✓

   The requested output is explanatory and knowledge-based, so retrieval-grounded generation matches the task.

Customers need grounded explanatory answers, not labels; use retrieval with generation for current catalog information.

7. Route paper questions to retrieval and arithmetic: Which architecture is most appropriate?

1. Route paper questions to retrieval and arithmetic questions to the calculator. ✓

   Intent-based routing selects the relevant capability for each request and avoids unnecessary calls.
2. Use retrieval for every request, then calculate only if retrieval returns no paper.

   Arithmetic requests still trigger an unnecessary retrieval call before reaching the calculator.
3. Send every request to the calculator, then use retrieval when the result is not numeric.

   Paper questions receive an unnecessary calculator call before retrieval, contrary to the routing requirement.
4. Run both tools for every request and ask the model to reconcile their outputs.

   Executing both tools regardless of intent wastes calls and does not meet the requirement to avoid unnecessary tool use.

Use intent-based routing so paper questions invoke retrieval and arithmetic questions invoke the calculator only when needed.

8. Check every required field and type: What is the next validation step?

1. Retrieve the source contract, populate any discovered values, and accept the record without rerunning complete schema validation.

   Authoritative retrieval may provide a missing value, but bypassing complete validation still fails to verify the final record’s required fields and types.
2. Accept the record because `vendor` passed validation.

   A single valid field does not make the complete record valid; the amount and required date still fail the contract.
3. Check every required field and type, then reject or route the incomplete record for review. ✓

   The amount requires numeric type handling, and the missing expiration date must cause rejection or review rather than unsupported completion.
4. Convert the amount and ask the model to estimate the missing date.

   The amount may need type conversion, but estimating a missing required date would invent data.

Validate all required fields and types; the string amount needs handling and the missing date requires rejection or review.

9. Use Supervisor Agent to route specialists: Which implementation best meets these constraints?

1. Use a coordinator that invokes every specialist, then lets the ticket specialist update records whenever it receives a request.

   Always invoking every specialist ignores conditional selection, and direct updates omit the required approval before consequential changes.
2. Use separate specialists with permission checks, but have an external workflow merge their outputs after each request and approve ticket changes automatically.

   Separate permissions help, but automatic approval violates the explicit human or control approval requirement for consequential ticket updates.
3. Use Supervisor Agent to route specialists, combine results, and require approval for ticket changes while preserving each tool’s permissions. ✓

   Supervisor Agent provides coordination, while specialist tool permissions and approval controls remain enforced at the appropriate boundaries.
4. Use a shared agent with all specialist tools and let it decide which actions to perform without approval.

   A shared unrestricted agent removes specialist permission boundaries and permits consequential ticket changes without the required approval.

Supervisor Agent coordinates specialists while preserving tool permissions and approval before consequential ticket changes.

10. p95 latency gate: Which gate blocks release?

1. p95 latency gate. ✓

   The measured p95 latency is 3.4 seconds, exceeding the required limit below two seconds.
2. Citation support gate.

   The measured citation support of 97% exceeds the required 95%; this gate passes.
3. Answer correctness gate.

   The measured correctness of 94% exceeds the required 90%; this gate passes.
4. Retrieval recall@5 gate.

   The measured retrieval recall of 91% exceeds the required 85%; this gate passes.

The latency gate fails; all three quality measures exceed their required thresholds.

11. Read the current record: Which implementation best prevents overwriting a newer change?

1. Generate an identifier from the customer name and submit the update without reading the existing record.

   A generated identifier does not establish the current record or version, so it cannot reliably detect a concurrent change.
2. Submit the update immediately and inspect the dashboard afterward for evidence of a conflict.

   Post-update inspection may reveal a problem but cannot prevent an update from overwriting a newer change.
3. Ask the language model to infer whether another process changed the dashboard since the request began.

   Model inference cannot replace an authoritative version check performed by the update operation.
4. Read the current record, then submit the change conditionally with its identifier and version. ✓

   Reading current state and requiring the update to match that version enables the operation to reject a concurrent newer change.

Read authoritative state and perform a version-conditional update that rejects mismatches.

12. Define the extraction fields: What prerequisite remains?

1. Define the extraction fields, allowed values, and output types before generating results. ✓

   Structured entity extraction requires an explicit schema describing fields, values, and expected types.
2. Add more retrieved documents so the model can generate a broader incident narrative.

   Additional context may help coverage but does not define the entities or constraints analysts require.
3. Increase response creativity so the tool can infer unusual incident categories.

   Creativity does not establish consistent field names, allowed values, or machine-readable output.
4. Replace extraction with a free-form summary and let analysts identify entities manually.

   Manual interpretation removes the consistent structured output required for triage workflows.

Define a structured extraction schema with fields, permissible values, and types before extracting entities.

Other Databricks GenAI Engineer domains

• Application Development — 292 questions →
• Assembling and Deploying Applications — 212 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 →