Databricks Generative AI Engineer Associate: how to prepare

Over half of the Databricks Generative AI Engineer Associate exam is about building and shipping an app. Application development is 30% of the questions and assembling and deploying another 22%. The rest is designing the app (14%), preparing its data (14%), evaluating and monitoring it (12%) and governing it (8%). That split decides how to prepare: build one retrieval app from start to finish, rather than read about each feature on its own.

The exam is 45 scored questions in 90 minutes, multiple choice and multiple selection, with up to 10 unscored ones the time already allows for. It costs $200 plus tax, is taken online with a proctor in English, Japanese, Brazilian Portuguese or Korean, and lasts 2 years. There is no prerequisite; Databricks recommends 6 months of hands-on work on the platform.

One project covers the whole guide

The 6 sections follow the life of an LLM application in order, so one project can walk through all of them. Take a support assistant that answers customers from your product manuals and cites the page. Step through it below: each section of the exam lights the part of the app it builds.

Try itOne RAG app, built in the order of the exam

1. 1Design
2. 2Data preparation
3. 3Application development
4. 4Assembling and deploying
5. 5Governance
6. 6Evaluation and monitoring

BackNext: Data preparation

1. Manuals
2. →Chunks table
3. →Search index
4. →Agent
5. →Endpoint
6. →Readers

What you build

Write down what goes in and what comes out: a customer's question in, a short answer with the manual page it came from out. Decide whether an Agent Bricks Knowledge Assistant does the job or the app needs its own chain, and which tools it calls in which order.

What the questions ask

Prompts that force a format, picking chain components for an input and output, and when Agent Bricks fits.

Build it once, slowly, in a workspace. Write down why you chose each setting as you go. The most common first word in the guide's 56 objectives is "select", 11 times: a model, a chunking strategy, a metric, each for a stated case. The reasons you wrote down are what those questions test.

Where people lose marks

Chunking is a decision, not a default

Data preparation asks you to pick a chunking strategy for a document's structure and the model's limits, and application development comes back to it from the other side: choose a chunking strategy from your model and retrieval evaluation. So the guide treats chunk size as something you measure and change. Try 2 strategies on your manuals, such as fixed-size chunks and chunks cut at section headings, and compare retrieval quality on the same 20 questions. Our chunking page explains the trade-offs, and the embedding model's context length sets the ceiling on how long a chunk can be.

Search settings follow the data

AI Search, the product the guide still calls Vector Search, finds nearest neighbours with HNSW and measures closeness by L2 distance, so you normalise embeddings if you want cosine similarity. Hybrid search adds a BM25 keyword search to the vector search and merges the two rankings, which matters for anything an embedding treats as noise: part numbers, error codes, SKUs. A Delta Sync Index keeps itself in step with the chunks table; know when you would compute the embeddings yourself instead. Objective 4.10 asks you to configure the index for a number of embeddings, an update rate, a latency and a cost, so practise those trade-offs, not only the setup.

Judges that need an answer key

Evaluation questions often turn on one detail: whether a judge needs ground truth. In MLflow's built-in judges, Correctness, RetrievalSufficiency and ToolCallCorrectness need the expected answer in the evaluation data. RelevanceToQuery, RetrievalRelevance, RetrievalGroundedness, Safety and Guidelines need only the question and the response. That decides what you can run on live traffic, where there is no answer key, and what needs a labelled evaluation set first. Our LLM-as-a-judge page explains how a judge works.

Batch inference is not a loop

For labelling a table of past tickets, the guide wants ai_query, and Databricks' own advice is to send the whole table in one query and let AI Functions handle parallelism and retries. Splitting the table into batches by hand reduces throughput. Setting failOnError to false keeps the rows that worked and returns an error for the ones that did not, instead of failing the run.

Whose permissions the app uses

An app on Databricks Apps runs by default as its own service principal, and every user shares its permissions. A row filter that limits what each manager sees then never applies, because the query does not run as the manager. User authorization makes the app act as the person asking, so Unity Catalog's row filters and column masks apply to them. This fits the governance section and deployment at the same time, so expect it in either.

The names changed after the guide

The guide was published on 18 March 2026, and several products have new names since. Vector Search became AI Search on 1 June 2026. Genie spaces became Genie Agents on 8 July 2026. The AI Gateway pages for model serving endpoints now sit under Unity Gateway in the documentation. The questions were written to the guide, so expect the older names there and the newer ones in everything you read. Our Databricks rename table lists each one with its date.

Two other names in the guide are worth knowing well. Agent Bricks covers Knowledge Assistant, Multiagent Supervisor and Information Extraction, and objective 1.6 asks when one of them solves the problem without a custom chain. The Prompt Registry keeps prompts in Unity Catalog, each version fixed, with an alias such as production that you move to a new version when you promote it. That is how the guide's prompt version control and promotion across environments work in practice.

Practice and timing

When the build is done, test yourself. Our free Generative AI Engineer Associate sample exam has 11 questions of our own on the 6 sections, with the reason behind every answer. Daily Bytes has a Generative AI Engineer Associate prep course, and its Android app has 3 full mock exams, timed and scored by section, with Daily Bytes Plus. A low score in one section tells you which part of the project to build again.

45 questions in 90 minutes is 2 minutes each, and multiple-selection questions take longer, so do not linger on the single-answer ones. Databricks does not publish the pass mark, and a failed attempt can be retaken after 14 days at the full fee.

If this is your first Databricks exam and you build pipelines as well as apps, our comparison of all 8 Databricks certifications shows where it fits. The Context Engineer Associate exam is the natural next one for agent builders.