Databricks GenAI Engineer practice test
Twenty questions across the six sections of the Databricks Generative AI Engineer Associate exam, weighted roughly as the real exam is. Application development carries the largest share, so it gets the most questions here.
Design Applications
Question 1. Stakeholders want an assistant that answers from 40,000 internal documents updated weekly. Which design should you propose?
- A. Fine-tune a model weekly on the documents
- B. Place all documents in the system prompt
- C. Retrieval-augmented generation over an index refreshed as documents change
- D. Train a model from scratch
Show answer
Answer: C
Retrieval-augmented generation keeps answers current as documents change, with no retraining. Fine-tuning would need repeating weekly and still would not guarantee factual recall.
Question 2. A requirement states answers must cite the source paragraph. What does this constrain in the design?
- A. Retrieval must return source identifiers and locations with each chunk
- B. The model must be the largest available
- C. Temperature must be set to one
- D. The application must run on GPUs
Show answer
Answer: A
Retrieval must return identifiers and locations alongside the text so citations can be rendered and verified. Model size and temperature do not produce provenance.
Question 3. Which requirement most strongly suggests an agent rather than a single-shot RAG call?
- A. Answers must be under 200 words
- B. Answers must cite sources
- C. The corpus is large
- D. The task requires choosing among tools and taking several dependent steps
Show answer
Answer: D
Needing to decide between multiple tools and take several dependent steps is the case for an agent. Simple lookup and summarisation do not need one.
Data Preparation
Question 4. Retrieval returns passages on the right topic that rarely contain the answer. Chunks are whole 30-page documents. What should change?
- A. Retrieve more chunks per query
- B. Reduce chunk size and add overlap between chunks
- C. Increase the model temperature
- D. Remove the system prompt
Show answer
Answer: B
Oversized chunks dilute embeddings so retrieval matches broad topic rather than specific content. Smaller chunks with overlap sharpen it.
Question 5. Documents contain tables whose meaning is lost when flattened to plain text. What is the appropriate preparation step?
- A. Preserve table structure or serialise tables in a form that keeps row and column relationships
- B. Delete all tables from the corpus
- C. Convert tables to images
- D. Increase chunk overlap only
Show answer
Answer: A
Preserving table structure, or converting tables to a textual representation that keeps row and column relationships, retains the meaning retrieval depends on.
Question 6. The corpus mixes public marketing content and confidential contracts. External users must never see contract content. What belongs in preparation?
- A. An instruction telling the model to avoid contracts
- B. A disclaimer on answers
- C. Separate indexes, or access metadata enforced at retrieval time
- D. Shorter chunks for contract documents
Show answer
Answer: C
Separating indexes or attaching access metadata enforced at retrieval prevents confidential content reaching unauthorised users. Prompt instructions are not access control.
Application Development
Question 7. What do embeddings enable in a retrieval application?
- A. Compression of the source documents
- B. Semantic similarity search over text represented as vectors
- C. Encryption of sensitive passages
- D. Translation between languages
Show answer
Answer: B
Embeddings place text in a vector space where semantic similarity becomes distance, which is what similarity search uses. They are not compression or encryption.
Question 8. A hybrid search combines vector similarity with keyword matching. What problem does the keyword component address?
- A. Exact matches on identifiers, codes and rare terms that semantic search misses
- B. Reducing the cost of embeddings
- C. Translating queries automatically
- D. Enforcing access control
Show answer
Answer: A
Exact identifiers, product codes and rare terms are often missed by pure semantic search, which keyword matching catches reliably.
Question 9. Answers must follow a fixed JSON structure for a downstream system. What is the most reliable approach?
- A. Ask the model in the prompt to return JSON
- B. Set temperature to zero and hope
- C. Parse whatever comes back with regular expressions
- D. Constrain output to a defined schema and validate before use
Show answer
Answer: D
Constraining the output with a schema and validating before use catches deviations deterministically. Asking politely in the prompt is not enforcement.
Question 10. An agent must call a function that charges a customer. What should the implementation require?
- A. Logging the call and notifying afterwards
- B. Restricting the agent to office hours
- C. Explicit confirmation before the call executes
- D. A longer system prompt
Show answer
Answer: C
Explicit confirmation before a consequential, irreversible action keeps a human in the decision. Logging records the charge after it happened.
Question 11. Retrieved web content contains hidden text instructing the model to ignore its rules. Which mitigation is most direct?
- A. Isolate retrieved content from the instruction channel and treat it as data
- B. Block the domain that hosted the content
- C. Reduce chunk size
- D. Increase the number of retrieved chunks
Show answer
Answer: A
Treating retrieved content as untrusted data, isolated from the instruction channel, removes the injection mechanism. Blocking one domain does not generalise.
Question 12. Latency is dominated by retrieving 50 chunks and passing all of them to the model. Quality is unchanged above 8 chunks. What should you do?
- A. Increase the model size to process them faster
- B. Retrieve fewer chunks, reranking to keep the most relevant
- C. Cache the answers and serve stale results
- D. Remove the citation requirement
Show answer
Answer: B
Reducing to the number that actually contributes cuts both latency and token cost with no quality loss, ideally with reranking to pick the best ones.
Assembling and Deploying Applications
Question 13. A model serving endpoint must handle a tenfold traffic spike at month end without paying for that capacity all month. What applies?
- A. Autoscaling with configured minimum and maximum capacity
- B. Provisioning permanently for peak
- C. Manual scaling when someone notices
- D. Rejecting requests above a fixed rate
Show answer
Answer: A
Autoscaling with appropriate bounds matches capacity to demand. Permanently provisioning peak capacity wastes money the rest of the month.
Question 14. A new prompt version must be compared against the current one on live traffic before full rollout. Which deployment pattern fits?
- A. Replace the current version immediately
- B. Deploy to development only and infer from there
- C. Split traffic between versions and compare measured outcomes
- D. Deploy at night when traffic is lowest
Show answer
Answer: C
Serving both versions with split traffic allows a measured comparison with quick rollback. Replacing outright removes the comparison and the safety net.
Question 15. An application artefact must be reproducible six months later, including the exact prompt, model version and retrieval configuration. What is required?
- A. A copy of the notebook
- B. Versioned registration of the full configuration: prompt, model version, retrieval settings and code
- C. A screenshot of the settings page
- D. The name of the engineer who built it
Show answer
Answer: B
Versioning and registering the full application configuration, not just the code, makes the behaviour reproducible. Code alone omits the prompt and model version.
Question 16. An endpoint must not be reachable from the public internet but must serve an internal application. What applies?
- A. A public endpoint with a strong token
- B. A public endpoint with an IP allow-list
- C. Disabling the endpoint outside office hours
- D. Private network access so the endpoint is not publicly reachable
Show answer
Answer: D
Private network access restricts reachability to internal callers. An IP allow-list still exposes a public endpoint.
Governance
Question 17. Which control best prevents an application returning data a given user is not entitled to see?
- A. Enforce entitlements at retrieval so unauthorised content never enters the context
- B. Ask the model to respect permissions in the system prompt
- C. Filter the generated answer for sensitive words
- D. Limit response length
Show answer
Answer: A
Enforcing entitlement at retrieval, before content reaches the model, prevents unauthorised content entering the context at all. Post-hoc filtering is unreliable.
Question 18. What should be logged to support investigation without creating a large new data risk?
- A. Nothing, to minimise risk
- B. Full prompts and completions indefinitely
- C. Metadata, source references and decisions, with sensitive content redacted
- D. Only the total request count
Show answer
Answer: C
Metadata, retrieved document references and decisions, with sensitive content redacted, supports investigation while limiting exposure. Full prompt retention creates a high-value target.
Evaluation and Monitoring
Question 19. Which metric indicates whether an answer is supported by the retrieved context?
- A. Relevance
- B. Groundedness
- C. Latency
- D. Token count
Show answer
Answer: B
Groundedness measures support by the provided sources. Relevance measures whether the answer addresses the question, and latency measures speed.
Question 20. Answer quality degrades three months after launch with no code change. What should monitoring have detected?
- A. Rising storage costs
- B. Increased log volume
- C. Higher CPU usage
- D. Drift in the question distribution or in the indexed corpus
Show answer
Answer: D
Shifts in the question distribution or in the underlying corpus change retrieval effectiveness over time, which drift monitoring surfaces before users complain.
How did you do?
Sixteen or more correct suggests you are close. Below fourteen, the section guides here are the fastest fix. Databricks does not publish a passing score for this exam, so treat specific figures quoted elsewhere with caution.