Databricks Certified Generative AI Engineer Associate Practice Exam

What the Databricks Certified Generative AI Engineer Associate exam covers

• Design Applications96 questions
• Data Preparation95 questions
• Application Development176 questions
• Assembling and Deploying Applications152 questions
• Governance84 questions
• Evaluation and Monitoring108 questions

Free Databricks Certified Generative AI Engineer Associate practice test questions

A sample of 10 questions with answers and explanations. Sign up free to practice all 711.

1. Question 1Application Development

   A retail support bot uses Mosaic AI Vector Search to fetch product return policies. Before sending the user's question to the LLM, the app must combine the retrieved passages with the original question so the model answers using only that context. Which prompt construction pattern accomplishes this?

   • AInsert the retrieved passages into the prompt as a context block alongside the questionCorrect
   • BReference the retrieved passages by their Unity Catalog volume path in the question text
   • CReplace the user's question with only the single top retrieved passage
   • DSend the user question alone and rely on the LLM's pretrained knowledge of the policy
   ✓ Correct answer: A

   The correct move is to concatenate the retrieved return-policy passages with the user's question inside the prompt itself, typically under a labeled context section. This is the core mechanic of prompt augmentation in a Mosaic AI Vector Search RAG pipeline: the LLM has no independent access to the vector index, so anything it should reason over must physically appear as tokens in the prompt it receives. For this retail bot, injecting the exact policy text lets the model answer from that grounded content instead of guessing. Referencing a Unity Catalog volume path only gives the model an address it cannot open, replacing the question with a single passage throws away the user's original intent, and skipping retrieval entirely invites outdated or fabricated answers. The takeaway: augmentation means placing the retrieved text itself into the prompt, not a pointer to it.

   Why the other options are wrong

   • BA Unity Catalog volume path is a file address, not readable content; the model cannot open or fetch it on its own.
   • CReplacing the question with only the top passage discards the user's actual intent and breaks the interaction.
   • DRelying on the LLM's pretrained knowledge skips retrieval entirely, risking outdated or wrong policy details.
2. Question 2Application DevelopmentSelect all that apply

   An engineer selecting a model from a hub for a healthcare-adjacent application notices the model card includes a disclaimer that the model is not validated for clinical or diagnostic use. Which TWO actions are appropriate given this disclaimer? (Choose TWO)

   • AIgnore the disclaimer, since model cards are only informal suggestions
   • BAssume the disclaimer only applies to competing organizations, and not to this project team
   • CFine-tune the model once to permanently remove the disclaimer's restriction
   • DAvoid using the model for diagnostic decision-making, consistent with the documented limitationCorrect
   • ETreat the disclaimer as a legitimate constraint when scoping the application's intended use caseCorrect
   ✓ Correct answer: D, E

   A model card's disclaimer that it is not validated for clinical or diagnostic use is a documented limitation, and the appropriate response has two parts. First, the team should avoid using the model for diagnostic decision-making, since that is precisely the use case the vendor has stated the model was not validated for and using it anyway would introduce unassessed risk. Second, the team should treat the disclaimer as a legitimate constraint when scoping the application, meaning the intended use case should be defined to stay within what the model has actually been validated for, rather than drifting into the disclaimed territory. Ignoring the disclaimer because model cards are only informal suggestions misunderstands their purpose, assuming the disclaimer only applies to other organizations misreads what the limitation is actually based on, the use case, not the user, and fine-tuning the model does not remove a vendor's documented limitation or the liability concerns behind it. Respecting a validated-use boundary is a matter of responsible scoping, not something a technical workaround can bypass.

   Why the other options are wrong

   • ATreating a documented safety disclaimer as merely an informal suggestion misunderstands its purpose and ignores a real, stated limitation on validated use.
   • BA disclaimer's scope is defined by the use case it addresses, not by which organization happens to be using the model.
   • CFine-tuning a model does not remove a vendor's documented validation limitation or the liability disclaimer that accompanies it.
3. Question 3Application Development

   A team notices their agent's quality metrics were strong at launch but wants ongoing visibility into whether real user traffic is still getting good answers weeks later. Which phase of the GenAI life cycle addresses this ongoing need?

   • AMonitoringCorrect
   • BEvaluation
   • CData preparation
   • DApplication design
   ✓ Correct answer: A

   Monitoring is the phase built for exactly this need: continuously observing an already-deployed system's real production traffic, often via inference tables and Lakehouse Monitoring, to catch quality drift or degraded performance weeks or months after a strong launch. Evaluation, by contrast, is a point-in-time, pre-deployment comparison, so it cannot by itself tell the team whether real users are still getting good answers today. Data preparation and application design are earlier lifecycle phases concerned with building the data pipeline and the initial architecture, not with ongoing production observability, so neither addresses this team's need for continued visibility.

   Why the other options are wrong

   • BEvaluation is a point-in-time, typically pre-deployment comparison of versions, not an ongoing production observability practice.
   • CData preparation concerns building and cleaning the underlying data pipeline earlier in the life cycle, not production monitoring.
   • DApplication design is the early architectural phase of the life cycle, unrelated to ongoing observation of live traffic.
4. Question 4Data Preparation

   A data engineer installs the pytesseract Python package but OCR calls fail with an error that the OCR engine cannot be found. What is the most likely cause?

   • AThe Delta table schema does not allow string columns, unrelated to a missing OCR engine binary
   • BThe Vector Search index has not been created yet
   • Cpytesseract wraps the Tesseract OCR engine, which must be separately installed on the systemCorrect
   • DThe Unity Catalog volume is set to read-only
   ✓ Correct answer: C

   Installing the pytesseract Python package alone does not install the underlying Tesseract executable it calls out to, so on many systems that binary must be installed separately, for example through the OS package manager, and be discoverable on the system path before OCR calls will succeed. Delta table schema constraints, a missing Vector Search index, and read-only volume permissions would each surface as entirely different, unrelated errors and have no connection to an OCR engine lookup failure. The takeaway: pytesseract errors that mention a missing engine point to the system-level Tesseract installation, not to the Python package itself.

   Why the other options are wrong

   • ADelta table schema constraints on string columns have no connection to an OCR engine lookup error raised by pytesseract.
   • BA missing Vector Search index would surface as a query-time error, not an OCR engine not-found error during text extraction.
   • DA read-only Unity Catalog volume would cause a write failure, not an engine-not-found error when calling pytesseract.
5. Question 5Assembling and Deploying Applications

   A team needs a simple chain that answers customer questions using only the current product catalog. Which sequence correctly assembles the chain?

   • ASend the question directly to the LLM, then retrieve catalog chunks afterward to append as a footnote
   • BRetrieve catalog chunks, send them straight to the LLM as the entire prompt, and ignore the original question
   • CInsert the question into the prompt template, call the LLM, and only retrieve catalog chunks if the answer looks wrong
   • DRetrieve the most relevant catalog chunks, insert them into the prompt template with the question, then call the LLMCorrect
   ✓ Correct answer: D

   Retrieval has to happen before the prompt is assembled, since the retrieved chunks are what gets substituted into the template alongside the user's original question, and the LLM is called only once that fully assembled prompt exists. Calling the LLM before retrieval and appending catalog chunks afterward as a footnote means the model already generated its answer with no grounding in current catalog data, defeating the purpose of RAG. Sending only retrieved chunks with no question at all gives the LLM nothing to actually answer. And calling the LLM first, treating retrieval as a fallback only if the answer looks wrong, still generates an initial ungrounded response and leaves grounding as an afterthought rather than the foundation of the answer.

   Why the other options are wrong

   • ACalling the LLM before retrieval means the answer is generated with no grounding at all, and appending chunks afterward does not let the model actually use them.
   • BSending only retrieved chunks with no question gives the LLM nothing concrete to answer about.
   • CCalling the LLM first and treating retrieval as only a fallback when the answer looks wrong still produces an initial ungrounded response.
6. Question 6Assembling and Deploying Applications

   A team receives embedding vectors in real time from an external streaming system that does not write to a Delta table, and needs to upsert and delete individual vectors on demand via API. Which Vector Search index type fits this requirement?

   • AA Delta Sync index
   • BA Model Serving batch endpoint
   • CA Genie space
   • DA direct-access indexCorrect
   ✓ Correct answer: D

   A direct-access index is managed entirely through explicit API calls, letting a team upsert or delete individual vectors on demand with no dependency on any backing Delta table, which is exactly what fits a scenario where embedding vectors arrive in real time from an external streaming system that never writes to a Delta table at all. A Delta Sync index, by contrast, requires and automatically syncs from a source Delta table, which simply does not exist in this scenario. A Model Serving batch endpoint is for scoring workloads over batches of data, not for storing or querying individual vectors. And a Genie space is a natural-language SQL interface for asking questions over tables, entirely unrelated to vector storage or streaming ingestion. The lesson: when there is no backing Delta table and vectors need direct, on-demand API control, a direct-access index is the right type.

   Why the other options are wrong

   • AA Delta Sync index requires a backing Delta table it automatically syncs from, which does not exist here.
   • BA batch serving endpoint is for scoring workloads, not for storing or querying vectors.
   • CA Genie space is a natural-language SQL interface, unrelated to vector storage.
7. Question 7Design Applications

   Output from an LLM call needs to be loaded into a Python dict by downstream code with no manual cleanup. Which prompt design habit best avoids parsing failures?

   • AAsk for raw JSON only, with no markdown code fences or surrounding explanation, matching a stated schemaCorrect
   • BAsk the model to wrap the JSON in a markdown code block so it is easy for a person to read
   • CAsk the model to append a short natural-language summary after the JSON for extra context
   • DAsk the model to bold the key names so the structure is visually clear in the response
   ✓ Correct answer: A

   A naive JSON parser expects the entire response to be valid JSON and nothing else; any markdown fence, bold markup, or trailing sentence introduces characters that break that parse unless someone adds extra stripping logic first. The requirement here is explicitly 'no manual cleanup,' so the prompt has to remove those extra characters at the source rather than downstream. Naming the schema keeps the keys consistent call to call. Takeaway: for code that parses model output directly, ask for exactly the payload the parser expects, nothing wrapped around it.

   Why the other options are wrong

   • BA markdown code block adds fence characters around the JSON that a strict parser will fail on unless stripped first.
   • CA trailing natural-language summary after the JSON breaks a parser that expects the JSON object to be the entire response.
   • DBold markup injects formatting characters into what needs to be clean, machine-readable JSON, breaking a naive parse.
8. Question 8Design Applications

   A content-moderation step must check a generated response for safety issues using a guardrail tool before it reaches the user. What is the correct placement of this check in the pipeline?

   • ARun the safety check before any response has been generated, so nothing exists yet to check
   • BDeliver the response to the user first, then run the safety check afterward as a formality
   • CGenerate the response and deliver it immediately, skipping any safety check
   • DGenerate the response, run a safety check on it, then deliver it only if the check passesCorrect
   ✓ Correct answer: D

   A guardrail-style safety check has two positioning requirements at once: it needs the generated response as its input, so generation must happen first, and it must act as a gate before that response reaches the user, so delivery cannot happen until the check passes. The key concept is that safety checks are not just another processing step but a control point in the pipeline that can block delivery entirely. In this scenario, running the check before anything has been generated has nothing to inspect, delivering the response and checking afterward defeats the purpose since the user has already seen it, and skipping the check altogether removes the safeguard the design explicitly calls for. Remember: a guardrail belongs strictly between generation and delivery, never before generation and never after delivery.

   Why the other options are wrong

   • AThere is no generated content to inspect before a response has actually been produced, so the check would have nothing to check.
   • BDelivering the response to the user before running the safety check means an unsafe response could already have been seen by the time the check runs.
   • CSkipping the safety check entirely removes the exact safeguard the pipeline design calls for before content reaches the user.
9. Question 9Evaluation and Monitoring

   A team evaluated a 70B model and a 7B model and found the 7B model meets their quality bar for a simple intent-classification task at a fraction of the cost. What Databricks cost-control practice does choosing the 7B model represent?

   • AProvisioned throughput reservation
   • BAI Gateway rate limiting
   • CResponse caching
   • DRight-sizing the model to the taskCorrect
   ✓ Correct answer: D

   Choosing the 7B model represents right-sizing the model to the task, because the team measured that a smaller, cheaper model still clears the quality bar for a simple intent-classification task, and selecting the smallest model that meets the bar is the definition of right-sizing rather than defaulting to the biggest available model. The concept being tested is that model cost scales with size, so overpaying for a 70B model when a 7B model performs equivalently for the specific task at hand is unnecessary spend. This scenario is specifically about model selection, not a capacity reservation, a rate-limiting rule, or a caching mechanism, which are all separate cost levers. Takeaway: pick the smallest model that clears your measured quality bar.

   Why the other options are wrong

   • AProvisioned throughput reservation is a capacity-sizing mechanism for a chosen model's serving endpoint, not the act of choosing which model to use.
   • BAI Gateway rate limiting caps request volume for a consumer; it has nothing to do with which model was selected for the task.
   • CResponse caching reduces redundant calls to whichever model is deployed; it is a separate lever from the decision of which model size to use.
10. Question 10Governance

    A telecom customer-care bot must mask phone numbers in retrieved call-log excerpts, but the masking should vary: a support rep sees the last four digits, a compliance auditor sees the full number, and a customer talking to the bot sees nothing. What mechanism natively supports this three-tier disclosure?

    • AThree separate physical copies of the call-log table
    • BA single AI Gateway rate limit rule configured with three tiers
    • CA column mask function with conditional logic keyed to group membershipCorrect
    • DThree Vector Search indexes built with different embedding models
    ✓ Correct answer: C

    A single column mask can contain conditional logic that checks the caller's group and returns a different value for each tier, full number for compliance, last four digits for support, nothing at all for the customer, all evaluated against the same underlying call-log table with no duplication needed. This flexible, multi-branch masking is exactly the key concept Unity Catalog column masks are designed to support. Maintaining three separate physical copies of the table is wasteful and hard to keep synchronized as data changes, a rate limit rule with tiers still only controls request volume rather than what data is shown, and building three separate Vector Search indexes with different embedding models addresses retrieval quality, not tiered data disclosure.

    Why the other options are wrong

    • AMaintaining three separate physical copies of the call-log table is wasteful and hard to keep synchronized as the underlying data changes.
    • BAn AI Gateway rate limit rule controls how many requests can be made; it has no mechanism for varying what data is disclosed.
    • DBuilding separate Vector Search indexes with different embedding models addresses retrieval quality, not tiered visibility of a specific field's value.

Who this Databricks Certified Generative AI Engineer Associate practice exam is for

This practice set is for anyone preparing for the Databricks Certified Generative AI Engineer Associate exam at the intermediate level - from first-time candidates building a foundation to experienced Data practitioners doing a final review before test day. If you learn best by working through realistic questions and reading why each answer is right or wrong, it is built for you.

How to use this Databricks Certified Generative AI Engineer Associate practice exam

1. Start with the free sample questions above to gauge your current baseline.
2. Read the full explanation on every question, including why each wrong option is wrong.
3. Track your weak domains and focus your study where you are losing the most marks.
4. Once you are scoring consistently well, take a timed, full-length mock exam.
5. Use your readiness score to decide when you are ready to book the real Databricks Certified Generative AI Engineer Associate exam.

Related Data resources

• Databricks Certified Generative AI Engineer Associate study guideKey concepts
• Data practice examsAll Data
• Certification pathWhere this fits
• Certification exam guides & tipsBlog
• Plans & pricingFree & paid
• How these questions are written and reviewedMethodology
• Report a problem with a questionCorrections
• Databricks Certified Machine Learning Associate practice examRelated
• MongoDB Associate Developer practice examRelated
• SnowPro Advanced practice examRelated

Databricks Certified Generative AI Engineer Associate practice exam FAQ