Databricks Generative AI Engineer Associate Practice Test

Answer choices

1. A. A query-rewrite step, retriever, prompt composer that retains the original question, generator, and MLflow tracing around the chain steps.
2. B. A generator plus a final regex parser; tracing is unnecessary when latency is the only metric.
3. C. A retriever only, because the embedding model can both rewrite and answer the question.
4. D. A multi-agent supervisor with no explicit retriever because subagents can infer the knowledge.

Wrong-answer review

• B. A generator plus a final regex parser; tracing is unnecessary when latency is the only metric.: A generator and parser omit evidence retrieval and the requested step-level latency visibility.
• C. A retriever only, because the embedding model can both rewrite and answer the question.: A retriever does not perform all of rewriting, response generation, and trace instrumentation.
• D. A multi-agent supervisor with no explicit retriever because subagents can infer the knowledge.: Orchestration does not replace a defined retrieval component or guarantee traceability.

Answer choices

1. A. Hard-code the latest numeric version in every deployment and rebuild the service for each promotion.
2. B. Edit version 17 in place for staging, then copy the text back if production needs rollback.
3. C. Load prompts by environment-specific aliases (for example staging/prod) that point to immutable prompt versions and move aliases only after evaluation gates pass.
4. D. Store the prompt only in the model source code because external prompt versioning prevents reproducibility.

Answer choices

1. A. Retrieve a runbook first, restart if any restart instruction is present, and inspect telemetry afterward.
2. B. Run telemetry and runbook retrieval in any order, but give the restart tool permission to self-authorize.
3. C. Read telemetry, retrieve the relevant runbook, verify the signature and authorized action, then invoke restart; otherwise escalate.
4. D. Restart whenever telemetry shows high latency because remediation should precede diagnosis.

Answer choices

1. A. Use user authorization for the background job even when no user is present and store the last user's token.
2. B. Use app authorization/service principal for the shared background telemetry and user authorization for interactive data operations that require the caller's permissions.
3. C. Use only the app service principal for all operations and reimplement row-level permissions in application code.
4. D. Use no OAuth and grant CAN MANAGE to all app users.

Answer choices

1. A. Training compute estimate only, because larger training runs imply stronger governance.
2. B. License/usage terms, supported deployment/data-handling characteristics, and relevant capability constraints.
3. C. Parameter count, marketing popularity, and benchmark leaderboard rank only.
4. D. Tokenizer vocabulary size only, because it determines legal usage.

Answer choices

1. A. Ask the generator to ignore offensive text but retrieve it normally.
2. B. Index the raw text unchanged because historical accuracy always requires exposing every word to the model.
3. C. Delete the entire source dataset even though the required factual fields are useful and can be separated.
4. D. Transform/extract the required structured facts while excluding or masking unnecessary harmful text from the retrieval corpus, preserving governed provenance to the source.

Answer choices

1. A. Generation fluency, because well-written answers can compensate for missing retrieval evidence.
2. B. Recall/coverage of required evidence in the retrieved candidate set, using labeled queries that identify the relevant source chunks.
3. C. Prompt length, because shorter prompts always improve retrieval recall.
4. D. Model-serving throughput, because more QPS increases the probability of finding relevant chunks.

Answer choices

1. A. Use model latency as the disclaimer compliance metric, assuming a response that meets the serving SLO also contains the required jurisdiction-specific wording.
2. B. Use a code-based custom scorer that deterministically checks the required disclaimer, alongside other semantic judges as needed.
3. C. Use RetrievalRelevance because a relevant document guarantees the output contains the required text.
4. D. Use a general safety judge only, because deterministic text requirements should never be coded.

Answer choices

1. A. Add every policy to the prompt so the model can choose which one applies.
2. B. Retrieve the generic FAQ corpus normally and add 'lost_card' as a weak keyword.
3. C. Route the lost-card intent to the approved lost-card context/policy and inject the required warning as trusted instruction, excluding incompatible generic troubleshooting content.
4. D. Let the model infer the safety warning from pretrained banking knowledge.

Answer choices

1. A. Increase the LLM context window so it can remember not to delete extra files.
2. B. Validate and constrain tool arguments against the user's authorized scope, use narrow schemas, and require confirmation for destructive/bulk operations.
3. C. Make the tool description longer so the model feels more certain.
4. D. Log deleted file names after execution and use the logs as the primary preventive control.