Databricks Generative AI Engineer Practice Questions

Below are 15 free practice questions for the Databricks Certified Generative AI Engineer Associate, each with the answer and a short explanation. They are sample questions written by HOW TO // AI to match the published exam guide. They are not real exam questions and not “dumps”; they’re original scenarios built from Databricks’ own objectives, and every answer can be checked against Databricks documentation.

The questions follow the exam’s section weights: 5 on Application Development, 3 on Assembling and Deploying Apps, 2 each on Design Applications, Data Preparation and Evaluation and Monitoring, and 1 on Governance.

Exam facts

From Databricks’ official exam page and exam guide, as of October 2026:

How to use these questions

• Answer before you read the explanation. Cover the answer line, commit to a letter, then check.
• Keep a steady pace. 45 scored questions in 90 minutes is about 2 minutes each, so give yourself 30 minutes for all 15.
• Tally by section. A miss in a 30% section costs more than a miss in an 8% one.
• Check the docs for every miss. Each explanation names the concept; look it up on docs.databricks.com and read the page.
• Build what you missed. Databricks Free Edition is enough to try most of these hands-on.

Two more things to know. Databricks doesn’t offer an official practice exam and says third-party practice results are often misleading, so treat any score here as a study signal, not a prediction. And the exam guide PDF has ten official sample questions with answers at the end; do those too.

Design Applications questions (14%)

Question 1 (Design Applications)

An HR team wants an internal chatbot that answers employee questions from about 400 policy PDFs. Answers must cite the source document, and HR subject matter experts want to improve quality by giving plain-language feedback. The team doesn’t want to write custom retrieval code. Which Agent Bricks option fits best?

• A. Information Extraction
• B. Knowledge Assistant
• C. Supervisor Agent (the multi-agent supervisor)
• D. A Genie Agent built on the PDF files

Answer: B. Knowledge Assistant builds a question-and-answer chatbot over your documents, returns answers with citations, and improves from natural-language feedback from subject matter experts. Information Extraction turns documents into a structured table, a Supervisor Agent coordinates several agents and tools, and Genie is built mainly for natural-language questions about data in tables, answered with generated SQL.

Question 2 (Design Applications)

A pipeline sends product reviews to a chat model on Databricks Foundation Model APIs and parses each reply as JSON with the fields sentiment, product and issue. Now and then the model adds a sentence before the JSON and the parser fails. Which change makes the endpoint return JSON that follows a schema you define?

• A. Set the temperature to 0
• B. Enable inference tables on the endpoint
• C. Pass a JSON schema in the request’s response_format field (structured outputs)
• D. Raise max_tokens so the full JSON always fits

Answer: C. Structured outputs let you request JSON that adheres to a specific schema through the response_format field, and Databricks recommends them for extraction and batch jobs that need a fixed format. A low temperature makes output less random but doesn’t enforce a format, and inference tables only log traffic.

Data Preparation questions (14%)

Question 3 (Data Preparation)

Your RAG source is a set of product manuals exported as Markdown, each with consistent section headers. With fixed 256-token chunks, answers often blend steps from two different procedures. Which change is most likely to help?

• A. Split on the Markdown headers and keep each section title as metadata
• B. Cut the chunk size to 32 tokens
• C. Remove all overlap between chunks
• D. Switch to a larger LLM

Answer: A. Databricks’ RAG data pipeline guide notes that fixed-size chunks rarely produce semantically coherent units, and recommends format-specific chunking for Markdown or HTML, for example splitting on headers. Adding metadata such as the section heading to each chunk can also improve retrieval. Tiny chunks lose context, and a bigger LLM can’t fix what retrieval hands it.

Question 4 (Data Preparation)

A retriever pulls the top 50 chunks from an AI Search index, but only the top 5 go into the prompt, and the most useful chunk often sits around rank 30. What does adding a reranker do here?

• A. Re-embeds the whole source table with a bigger model
• B. Compresses chunks so more fit in the context window
• C. Removes the need for an embedding model
• D. Re-scores the retrieved candidates for relevance to the query, so the most relevant ones move to the top

Answer: D. Reranking evaluates the documents a search already retrieved and reorders them by relevance, so the best matches are the ones passed to the LLM. Databricks says it adds a small latency cost but can significantly improve retrieval quality, and recommends trying it for RAG agents. In AI Search you can also name metadata columns for the reranker to consider.

Application Development questions (30%)

Question 5 (Application Development)

A customer-facing assistant runs on a Foundation Model APIs pay-per-token endpoint. Security asks you to block unsafe or harmful content in both user prompts and model responses without rewriting the app. What should you configure?

• A. Inference tables on the endpoint
• B. AI Guardrails on the endpoint through AI Gateway
• C. A per-user rate limit
• D. A higher temperature

Answer: B. AI Guardrails detect unsafe and harmful content in model inputs and outputs and stop the model from engaging with it, and they’re configured on the serving endpoint rather than in app code. Inference tables only record requests and responses, and rate limits only control volume. In newer Databricks docs, the same job is handled by guardrail service policies in Unity Gateway.

Question 6 (Application Development)

Your agent chains a retriever, two tool calls and an LLM. Some answers are wrong, and you need to see, for each request, what the retriever returned, which tools ran and how long each step took. Which feature gives you that?

• A. MLflow Tracing
• B. Unity Catalog lineage
• C. An MLflow model signature
• D. An endpoint rate limit

Answer: A. MLflow Tracing records the inputs, outputs, latency, token usage and cost of every intermediate step in an agent, so you can follow a single request from retrieval to final answer. Lineage shows how data assets relate, and a model signature describes input and output schemas, not what happened during a request.

Question 7 (Application Development)

You’re building a Supervisor Agent. One subagent answers policy questions from documents. Users also ask things like “What were last quarter’s returns by region?”, which need SQL over curated sales tables in Unity Catalog. What should you add as a second subagent?

• A. An AI Search index built from the sales table rows
• B. A fine-tuned LLM trained on a CSV export of the sales tables
• C. A Genie Agent (the exam guide calls these Genie Spaces) configured on the sales tables
• D. A system prompt that includes the full sales tables

Answer: C. A Supervisor Agent can coordinate Genie Agents, agent endpoints, Unity Catalog functions and MCP servers. Genie turns natural-language questions into SQL over governed Unity Catalog data, which is what an aggregate question needs. Vector search over table rows or a fine-tuned model can’t reliably compute totals from current data.

Question 8 (Application Development)

Which statement correctly contrasts evaluation and production monitoring for a generative AI app built with MLflow 3 on Databricks?

• A. Evaluation and monitoring need different scorers that can’t be shared
• B. Monitoring only works if you provide a ground-truth answer for every production request
• C. Evaluation can only run after the app is deployed
• D. Evaluation runs scorers on a dataset during development; monitoring runs scorers on a sample of live traces, and you can use the same scorers for both

Answer: D. Databricks’ docs say you can use the same scorer for evaluation in development and monitoring in production. Evaluation runs with mlflow.genai.evaluate() on a dataset, while production monitoring schedules scorers against a configurable sample of incoming traces from an app instrumented with MLflow Tracing.

Question 9 (Application Development)

Your chunks average about 900 tokens, but the embedding model you picked accepts a maximum of 512 tokens. What happens, and what should you do?

• A. The model splits each chunk into two embeddings automatically, so nothing changes
• B. Text past the limit is truncated, so part of each chunk is never embedded; choose an embedding model with a longer context or make the chunks smaller
• C. The index refuses to build until you switch to a different endpoint type
• D. Nothing, because embedding models have no input limit

Answer: B. Databricks’ RAG guidance says chunks that exceed an embedding model’s maximum token limit are truncated, which can lose important information, and gives bge-large-en-v1.5 and its 512-token limit as an example. Match the embedding model’s context length to your chunk size, or shrink the chunks.

Assembling and Deploying Apps questions (22%)

Question 10 (Assembling and Deploying Apps)

You try to create a Delta Sync index on a standard AI Search endpoint from a Delta table of chunks, and it fails a requirement check. The table doesn’t have row tracking enabled. What does the source table need?

• A. Change data feed enabled
• B. Partitioning by date
• C. Conversion to CSV
• D. A Direct Vector Access index in front of it

Answer: A. For standard endpoints, Databricks requires the source table of a Delta Sync index to use a change data feed. Tables with row tracking enabled use a change data feed automatically, which is why the question rules that out. Direct Vector Access is a different index type that you update yourself through the API.

Question 11 (Assembling and Deploying Apps)

Every night you need an LLM-written summary for about 2 million new support tickets stored in a Delta table. Which approach follows Databricks’ guidance for this batch workload?

• A. Loop over the rows in Python and call a real-time endpoint one request at a time
• B. Paste the tickets into AI Playground in batches of 50
• C. Run ai_query() over the full table in a scheduled job, using a Databricks-hosted foundation model
• D. Fine-tune a new model every night on the day’s tickets

Answer: C. ai_query() supports batch inference, and AI Functions handle parallelization, retries and scaling for you. Databricks advises submitting the full dataset in a single query rather than splitting it into small batches, and using Databricks-hosted foundation models (the ones prefixed databricks-) for batch work.

Question 12 (Assembling and Deploying Apps)

You’ve logged a RAG chain with MLflow 3 and need to register it in Unity Catalog so it can be governed and served. How do you name it?

• A. With a one-part name such as rag_chain; Unity Catalog adds the rest
• B. With the notebook path where the chain was built
• C. With the endpoint name you plan to serve it on
• D. With a three-level name such as main.support.rag_chain (catalog.schema.model)

Answer: D. Models in Unity Catalog use a three-level name in the form <catalog>.<schema>.<model>. In MLflow 3 the default registry URI is already databricks-uc; with older clients you set it with mlflow.set_registry_uri("databricks-uc") before registering.

Governance question (8%)

Question 13 (Governance)

Support tickets that feed your RAG index contain customer names, email addresses and phone numbers. You want to replace those values in a SQL step before indexing. Which function does that?

• A. ai_classify()
• B. ai_mask(content, array('person', 'email', 'phone'))
• C. ai_query() with temperature set to 0
• D. vector_search()

Answer: B. ai_mask() masks the entity types you list, such as person, email, phone or address, and returns the text with those values replaced. ai_classify() assigns labels to text, and vector_search() queries an index. Note that ai_mask() was in Public Preview when we checked the docs.

Evaluation and Monitoring questions (12%)

Question 14 (Evaluation and Monitoring)

You have 200 real user questions but no reference answers yet. Which built-in MLflow judge can’t run until you add ground truth?

• A. Safety
• B. RelevanceToQuery
• C. Correctness
• D. RetrievalGroundedness

Answer: C. Correctness checks the response against ground truth you provide, so every question needs expectations (the reference answer or facts) before it can run. Safety, RelevanceToQuery and RetrievalGroundedness judge the inputs, outputs and retrieved context without a reference answer.

Question 15 (Evaluation and Monitoring)

Costs on a pay-per-token endpoint jumped after one analyst started running large scripts against it. You want to cap that person’s usage without blocking the rest of the team. What should you configure?

• A. A custom per-user rate limit, in queries or tokens per minute, on the endpoint
• B. An endpoint-wide rate limit set very low
• C. Turning off inference tables
• D. Switching the endpoint to a larger model

Answer: A. AI Gateway rate limits can be set per endpoint, as a default per user, or as custom limits for specific users, service principals or groups, measured in queries per minute or tokens per minute. A custom user limit overrides the default, while an endpoint-wide limit caps everyone and blocks all requests once it’s exceeded.

How did you do?

• 13 to 15 correct: your concepts are solid. Spend your remaining time on hands-on builds and re-read the exam guide line by line.
• 9 to 12 correct: close. List the sections you missed and work through them in our 6-week Databricks study plan.
• 8 or fewer: start with the study plan from week 1 and build the hands-on checklist before booking.

Watch the names while you study. The exam guide still says Vector Search, Genie Spaces and Multiagent Supervisor, while the current docs say AI Search, Genie Agents and Supervisor Agent. Databricks calls the exam guide the source of truth for the exam, so learn both. If new terms like RAG, embeddings or guardrails still feel fuzzy, our plain-English AI glossary covers them.

Generate more practice with AI

You can have a chatbot write extra questions, as long as you make it cite its sources. Upload the exam guide PDF and paste this:

Act as an exam coach for the Databricks Certified Generative AI Engineer Associate. The attached exam guide is the only source of objectives.

Write 5 new scenario-based multiple-choice questions for the objective: [PASTE ONE OBJECTIVE FROM THE GUIDE].

Rules:
- 4 options each, exactly one correct, with plausible wrong answers.
- Use current Databricks product names and mention the older name if the guide uses one.
- Show one question at a time and wait for my answer.
- After I answer, explain why the right option is right and why each wrong one is wrong, and cite the docs.databricks.com page you relied on.
- If you can't verify a detail in the official docs, say so instead of guessing.

Act as an exam coach for the Databricks Certified Generative AI Engineer Associate. The attached exam guide is the only source of objectives.

Write 5 new scenario-based multiple-choice questions for the objective: [PASTE ONE OBJECTIVE FROM THE GUIDE].

Rules:
- 4 options each, exactly one correct, with plausible wrong answers.
- Use current Databricks product names and mention the older name if the guide uses one.
- Show one question at a time and wait for my answer.
- After I answer, explain why the right option is right and why each wrong one is wrong, and cite the docs.databricks.com page you relied on.
- If you can't verify a detail in the official docs, say so instead of guessing.

AI-written questions can be confidently wrong, so check any surprise against the docs. Our guide to making AI more accurate and honest shows how to catch it.

Try it now

Set a 30-minute timer, redo every question you missed without looking at the explanation, then open the docs page for each one. If you’re still deciding whether to book, check the Databricks exam cost first, or compare it with other options in our ranking of AI certifications.

Related guides

• Databricks Generative AI Engineer Exam Cost
• Free AI exam prep for other certifications
• The Best AI Certifications in 2026, Ranked: Which Should You Get?
• Free NVIDIA NCA-GENL Practice Questions
• Free AWS Machine Learning Engineer – Associate Practice Questions

HOW TO // AI is not affiliated with or endorsed by Databricks. These are original practice questions, not questions from the real exam. Exam details come from Databricks’ official certification pages and exam guide as of October 2026 and can change. Check the official exam page before you book.