Are these real Databricks GenAI exam questions?

No — and you should be wary of any site that says its questions are. Real exam items are confidential. These questions are written from scratch against the published Databricks GenAI exam guide to match its style, difficulty and domain weighting.

Should I use Databricks GenAI dumps?

No. A dump is a collection of leaked or memorised real exam items, and using one breaks the Databricks certification agreement — the vendor can revoke your certification and ban you from future exams. Dumps also tend to fail on their own terms: their answer keys are unverified, they drift out of date as the exam changes, and memorising answers leaves you stuck the moment a question is reworded. Original practice questions with an explanation on every option teach the reasoning the Databricks GenAI actually tests, which is what carries over to questions you have never seen.

Do the Databricks GenAI questions come with explanations?

Every option carries its own explanation — not just the correct one. You find out why the answer you nearly picked was wrong, which is where most of the learning actually happens. Each question also links back to the revision lesson that teaches the concept.

How many practice questions should I do before the Databricks GenAI exam?

Volume matters less than what you do with the misses. A useful signal is consistency rather than a total: when you can hold a comfortable margin above the 700 pass mark on questions you have never seen before, across every domain rather than just your strong ones, you are ready.

Can I retry the questions I got wrong?

Yes. Wrong answers and flagged questions go into a review pool you can drill on their own, so you can grind exactly the material that is costing you marks instead of re-answering the ones you already know.

Databricks GenAI practice

Free Databricks GenAI practice questions

Drill exam-realistic Databricks Certified Generative AI Engineer Associate questions by domain, with an explanation on every option — not just the right one. 5 fully worked examples are further down this page, answers included.

Question bank: 599
across 6 domains
Free, no account: 5/day
a free account allows 30 a day
Real exam: 45 Qs
90 min · pass Pass / fail
Explanations: Every option
right and wrong

Build a practice session

5 free questions left today.

Domains

How many?

Mode

Ready when you are

10 fresh questions drawn across all domains, in Learn mode.

Focused review

Every question you answer incorrectly, and every question you flag while practising, is saved here automatically. Finish a session and you can come back to re-drill just those.

5 sample Databricks GenAI questions, fully explained

Real questions from the Databricks GenAI bank, with the answer key and the reasoning behind every option. Read them, then go back up the page and try the rest.

Question 1Design Applications

A Generative AI Engineer at a telecom company must place each incoming support ticket into one of five existing queues: billing, outage, device, porting or cancellation. Each ticket goes to exactly one queue, and the queue name is written back to the ticketing system. Which model task should the engineer select for this requirement?

Choose one.

a
Text summarization of each ticket into a short queue description
A summary condenses the ticket but returns free text, not one of the five queue names the ticketing system expects.
b
Named entity recognition over each ticket to find the queue terms
Entity recognition pulls spans such as phone numbers or plan names out of the text; it does not decide which queue the ticket belongs to.
c
Text classification of each ticket against the five queue names Correct
Routing a ticket to exactly one of a fixed set of known queues is single-label text classification; the output is one of the five labels.
d
Text generation that drafts a reply to each ticket for the queue
Drafting a reply is a later step for an agent; it does not produce the routing decision the requirement asks for.

The concept

Model task selection starts from the output the business needs: a choice among known categories is classification.

Why that’s the answer

The required output is one value from a closed list of five queues, and each ticket gets exactly one. That is single-label text classification. Summarization and generation return free text that would still need to be mapped to a queue, and entity recognition extracts spans from the ticket rather than deciding a category.

How to reason it out

Look at the output the system needs: one of five fixed queue names.
A fixed, known set of categories means classification.
Exactly one queue per ticket means single-label, not multi-label.

Exam tip: Closed list of categories as the output: choose text classification.

Databricks GenAI: Prompt Design, Output Formats and Model Task Selection — the lesson that teaches this.

Question 2Data Preparation

A Generative AI Engineer is building a RAG assistant over 4,000 signed supplier contracts that were scanned to PDF years ago. A test extraction with pypdf returns empty strings for almost every page, because the files contain only page images with no text layer. Which approach should the engineer use to get the contract text?

Choose one.

a
Open each file with PyMuPDF and call get_text() on each of its page objects
PyMuPDF's get_text() reads the PDF's text layer, just as pypdf does. With image-only pages there is no text layer, so it also returns almost nothing.
b
Render pages to images and run Tesseract OCR on them through pytesseract Correct
The pages are images, so the text has to be recognised optically. pytesseract wraps the Tesseract OCR engine and returns the text that appears in each rendered page image.
c
Use pdfplumber to collect the words and characters found on each page
pdfplumber is excellent for text-based PDFs and tables, but it works from the embedded character objects. Scanned pages have none, so it cannot recover the text.
d
Call pypdf extract_text() with extraction_mode set to "layout"
Layout mode only changes how pypdf arranges text it finds in the text layer. It adds no OCR, so image-only pages still come back empty.

The concept

Scanned PDFs are images: extracting their text needs OCR, not a text-layer PDF parser.

Why that’s the answer

The symptom (empty strings from pypdf) tells you the PDFs have no text layer. pypdf, PyMuPDF get_text() and pdfplumber all read that embedded layer, so switching between them changes nothing. Only an OCR engine such as Tesseract (via pytesseract) can turn page images into text.

How to reason it out

Read the symptom: empty extraction means no text layer, not a parser bug.
Separate text-layer parsers (pypdf, PyMuPDF, pdfplumber) from OCR engines.
Choose OCR (pytesseract over rendered page images) for scanned documents.

Exam tip: No text layer means OCR: text-layer PDF parsers cannot read scanned pages.

Databricks GenAI Engineer: Document Extraction, Filtering and Chunking into Delta Tables — the lesson that teaches this.

Question 3Application Development

A Generative AI Engineer is building a document Q&A app on Databricks. The flow is fixed and linear: retrieve chunks from a Vector Search index, fill a prompt template, call a chat model and parse the reply into a field. The engineer wants ready-made, swappable components for each of those steps instead of writing glue code. Which tool fits best?

Choose one.

a
DSPy, compiling the prompt template steps against a metric and labeled examples
DSPy optimizes prompts and modules against a metric. The scenario has no metric or training set and asks only for ready-made components to wire together.
b
Hugging Face Transformers, running the chat model in a text-generation pipeline
Transformers loads and runs models. It offers no retriever, prompt template or output-parser components to assemble a RAG flow.
c
MLflow Tracing, recording the retriever, prompt and chat model calls as spans
Tracing observes an application that already exists; it does not supply the components that make up the chain.
d
LangChain, chaining a retriever, prompt template, chat model and output parser Correct
LangChain provides exactly these building blocks and composes them into a chain, with Databricks integrations for the retriever and the chat model.

The concept

LangChain is the general-purpose framework for composing LLM application components into chains.

Why that’s the answer

The requirement is a fixed sequence of standard steps with prebuilt, interchangeable parts. That is LangChain's core purpose: retrievers, prompt templates, chat models and output parsers composed into one chain. DSPy targets metric-driven optimization, Transformers is a model runtime, and MLflow Tracing only observes what runs.

How to reason it out

Note the shape: a fixed, linear flow of standard steps.
Ask which tool supplies components for every step.
Rule out tools that optimize, run models or observe instead of compose.

Exam tip: Fixed pipeline of standard steps: compose it with LangChain components.

Databricks GenAI Engineer: LangChain, Prompt Augmentation and Prompt Engineering — the lesson that teaches this.

Question 4Assembling and Deploying Applications

A Generative AI Engineer works in a workspace whose default catalog is hive_metastore, on a cluster with MLflow 2.x. A pyfunc RAG chain has been logged to a run, and it must now be registered as a governed model under the catalog ml and schema support. Calls to register it keep creating models in the legacy workspace registry. Which line should the engineer run before registering?

Choose one.

a
mlflow.set_tracking_uri("databricks-uc")
The tracking URI controls where runs and experiments are logged, not which model registry receives registrations.
b
mlflow.set_registry_uri("databricks-uc") Correct
This points the MLflow client's model registry at Unity Catalog, so the next registration creates a Unity Catalog model.
c
mlflow.set_registry_uri("databricks")
The databricks registry URI is the legacy workspace model registry, which is exactly what the engineer is trying to leave.
d
mlflow.set_tracking_uri("databricks")
The tracking URI only decides where runs and experiments are logged, and databricks is already the workspace default; registrations still go to the legacy registry.

The concept

The MLflow registry URI decides where registered models are created; databricks-uc means Unity Catalog.

Why that’s the answer

With MLflow 2.x and a non-Unity Catalog default catalog, the client still registers to the workspace registry. Setting the registry URI to databricks-uc switches registration to Unity Catalog. Setting the tracking URI, to either value, is a near-twin that only moves runs and experiments, and the plain databricks registry URI is the legacy registry. MLflow 3 uses databricks-uc by default, which is why the stem fixes the version.

How to reason it out

Separate the tracking target (runs) from the registry target (models).
Unity Catalog registration needs the registry URI databricks-uc.
Run it before calling register_model or log_model with a registered model name.

Exam tip: Registry URI databricks-uc = Unity Catalog models; the tracking URI is a different setting.

Databricks GenAI Engineer: Coding pyfunc Chains and Registering Them to Unity Catalog — the lesson that teaches this.

Question 5Governance

A Generative AI Engineer runs a customer-support chat model behind a Databricks Model Serving endpoint. Customers often paste their phone numbers and home addresses into questions. Compliance says those personal details must not reach the model, and the product owner requires such questions to still be answered rather than rejected. Which endpoint setting meets both requirements?

Choose one.

a
Input PII detection set to Block in the endpoint's AI Gateway
Block keeps the details from the model, but it rejects the whole request, which breaks the requirement to answer it.
b
The Safety guardrail enabled for the endpoint's request inputs
The Safety guardrail screens harmful content; phone numbers and addresses are not what it looks for.
c
Input PII detection set to Mask under the endpoint's AI Gateway Correct
Mask replaces the detected personal details and lets the request continue, so the model never sees them and the customer still gets an answer.
d
Per-user rate limits configured in the endpoint's AI Gateway
Rate limits cap request volume per user; they do nothing about personal details inside a request.

The concept

AI Gateway PII detection on a serving endpoint has three behaviours: Block, Mask and None.

Why that’s the answer

Both Block and Mask stop detected personal data from reaching the model, but Block rejects the request while Mask removes the details and lets the request proceed. When the objective is to keep answering, Mask is the setting. Safety screening and rate limits are other gateway features that do not detect personal data.

How to reason it out

Requirement 1: personal details must not reach the model, so PII detection must act on inputs.
Requirement 2: the request must still be answered, which rules out rejecting it.
Mask removes the details and forwards the rest; Block rejects the request.

Exam tip: PII must not reach the model and the request must still succeed = PII detection set to Mask.

Databricks GenAI Engineer: Guardrails, PII Masking and Data Licensing — the lesson that teaches this.

The Databricks GenAI question bank, by domain

The bank is built to the exam's own weighting, so the practice you get reflects the marks that are actually on offer — not whichever domain was easiest to write questions for.

Published Databricks GenAI practice questions per exam domain
Domain	Exam weight	Topics	Questions
Design Applications	14%	2	84
Data Preparation	14%	2	84
Application Development	30%	4	179
Assembling and Deploying Applications	22%	4	132
Governance	8%	1	48
Evaluation and Monitoring	12%	2	72
Total	100%	15	599

Practise Databricks GenAI one domain at a time

How Databricks GenAI questions are worded

Most Databricks GenAI questions are not asking whether you can recall a definition. They describe a situation and ask which option satisfies it — so the skill being tested is reading the requirement precisely and eliminating options that fail it.

Single-response vs multiple-response

A single-response question has exactly one right answer. A multiple-response question tells you how many to pick ("Choose TWO") and there is no partial credit — getting one of the two right scores nothing. Read that instruction before you read the options.

Read the last line first

The final sentence is the actual question; everything before it is scenario. Read it first, then read the scenario knowing what you are looking for. It stops you from building an answer in your head that the question never asked for.

Hunt for the qualifier

Most scenarios turn on one word — MOST cost-effective, LEAST operational overhead, with the LEAST latency, without changing application code. Two options are frequently both technically correct, and the qualifier is the only thing separating them.

Eliminate, then choose

Distractors are almost always real Databricks services doing a real job — just not this job. Rule out the ones that break a stated constraint before you compare what's left. On a question you truly don't know, eliminating two options turns a guess into a coin flip.

Practise Databricks GenAI from your browser

The SaveMyCert Chrome extension opens a daily drill of 5 Databricks GenAI questions in the side panel, with the full explanation behind every answer. No account needed.

Get the Chrome extension

Beyond Databricks GenAI practice

Databricks GenAI practice questions: your questions

Yes. You can answer 5 questions a day with no account at all, and a free account raises that to 30 a day across the full 599-question Databricks GenAI bank, with an explanation on every option. Pro removes the daily cap entirely.

Source

Exam structure, domain weights and scoring on this page come from the official Databricks GenAI exam guide.

Databricks Certified Generative AI Engineer Associate exam page and guide — Databricks