On this page
The short answer
- The real exam: 45 questions in 90 min, for $200.
- 5 sample questions below, from 5 of the exam’s 6 domains, each with its answer and why it’s right.
- They are original questions, written from Databricks’ published outline. They aren’t copies of real exam questions, and we don’t publish those.
- For a score by domain, take the free 10-question practice test. Nothing to download: it runs in your browser.
The Databricks Generative AI Engineer Associate exam at a glance
From our Databricks Generative AI Engineer Associate exam guide, which follows Databricks’ own page and is checked every week. Databricks hasn’t announced a retirement date for Generative AI Engineer Associate.
| Databricks Generative AI Engineer Associate | |
|---|---|
| Level | Associate |
| Written for | Generative AI Engineer |
| Questions | 45 |
| Time | 90 min |
| Passing score | Not published by Databricks |
| Cost | $200 |
| Recommended experience | 6+ months building generative AI solutions |
What the exam covers
The exam has 6 domains: Design Applications (14%), Data Preparation (14%), Application Development (30%), Assembling and Deploying Apps (22%), Governance (8%), Evaluation and Monitoring (12%). Application Development is the largest, at 30%, so it deserves the largest share of your study time.
The sample questions below start with one from each domain in turn, so you can see how each is asked.
5 sample questions, with answers
Answer each one before you open its answer. If you get it right for the wrong reason, count it as wrong: the explanation matters more than the letter.
Question 1 · Design Applications
A support app must answer from 10,000 product manuals that change every month. Which design fits?
- AFine-tune a model every month only
- BPaste all manuals into each prompt
- CRAG over a vector index of the manuals
- DAnswer from the base model alone
Show the answer
C. RAG over a vector index of the manuals
RAG retrieves the relevant manual sections at query time, so answers stay current without retraining. Monthly fine-tuning is costly, whole manuals overflow the context, and the base model doesn’t know them.
Question 2 · Data Preparation
Long PDFs must be prepared for retrieval. What should happen before embedding?
- AEmbed each whole PDF as one vector
- BDelete all headings and tables
- CTranslate them into one language only
- DParse and chunk them with some overlap
Show the answer
D. Parse and chunk them with some overlap
Parsing extracts the text, and chunking with overlap creates passages small enough for focused retrieval. One vector per PDF blurs meaning, and removing structure loses context.
Question 3 · Application Development
Which Databricks service stores embeddings and serves similarity search for a RAG app?
- ADelta Sharing recipients
- BDatabricks AI Search
- CDatabricks SQL alerts
- DLakeflow Jobs scheduling
Show the answer
B. Databricks AI Search
AI Search (formerly Vector Search) indexes embeddings, often synced from a Delta table, and serves similarity queries to RAG apps. Delta Sharing shares data, SQL alerts notify on query results, and Lakeflow Jobs schedule work.
Question 4 · Assembling and Deploying Apps
How should a RAG chain be packaged so it can be versioned and deployed to a serving endpoint?
- ASave it as a dashboard
- BPaste it into a SQL query
- CEmail the notebook to admins
- DLog it as an MLflow model
Show the answer
D. Log it as an MLflow model
Logging the chain with MLflow records its code, dependencies and signature, so it can be registered in Unity Catalog and deployed to Model Serving. Dashboards, queries and emails don’t package models.
Question 5 · Governance
Prompts and responses must be screened for unsafe content and personal data at the serving layer. What should be configured?
- AA bigger cluster for the endpoint
- BA comment on the source table
- CAI Gateway guardrails on the endpoint
- DA dashboard refresh schedule
Show the answer
C. AI Gateway guardrails on the endpoint
AI Gateway adds guardrails, such as safety filtering and PII detection, plus usage tracking and rate limits, to serving endpoints. Cluster size, comments and schedules don’t filter content.
How to use practice questions
Test before you study. A short test at the start shows which domains you already know, so you don’t spend a week on them.
Read every explanation, including the ones for questions you got right. Most exam questions turn on one detail in the scenario, and the explanation names it.
Then go full length. A few questions check your understanding; they don’t tell you whether you can hold your attention for 90 min. Before you book, sit full, timed practice exams until you pass them comfortably more than once.
Practice exams, downloads and dumps
You don’t need to download anything to practice. Our practice test is free and runs in the browser, on a phone or a computer.
Dumps are a different thing: collections sold or shared as the “real” exam questions. We don’t publish or link to them. They are taken from the exam in breach of the agreement every candidate accepts, they go out of date when Databricks refreshes its questions, and nobody stands behind their answers.
There is also a plainer reason. An exam is a sample of what the job needs. If you memorize the sample, you pass without the knowledge, and the first week of work shows it.
What to do next
Take the practice test, note your weakest domain, and build your study time around it.