The short version
The exam has six domains. Application Development (30%) and Assembling and Deploying Applications (22%) make up just over half of it. Weightings and topics here come from the exam guide dated March 18, 2026. The “typical traps” and practice ideas are our own advice, not vendor statements.
Domain 1: Design Applications (14%)
What it covers. Turning a business goal into an AI pipeline. You describe the inputs and outputs, write a prompt that elicits a specific format, choose model tasks and chain components, and order tools for multi-step reasoning. You also decide when to use Agent Bricks: Knowledge Assistant, Multiagent Supervisor, or Information Extraction.
In practice. A team wants an assistant over manuals. Do they need a hand-built agent with their own chunking, or is a managed starting point enough? Is the task question answering, extraction, or routing between specialists?
Typical traps. Choosing the most powerful option when a simpler managed one meets the need, and mixing up neighboring Agent Bricks types.
What to practice. Write one sentence for each Agent Bricks type saying what goes in and what comes out. Then write a prompt that returns JSON in a fixed shape.
Domain 2: Data Preparation (14%)
What it covers. Making documents ready for retrieval. You pick a chunking strategy for a document structure and model limits, filter out content that degrades quality, choose a Python package to extract text from a given format, and write chunks to Delta tables in Unity Catalog. The outline also includes identifying the source documents a RAG application needs, evaluating retrieval with tools and metrics, advanced chunking, and re-ranking.
In practice. Scanned images need text extraction; HTML pages need their navigation stripped; long procedures may need larger context than a small chunk gives you.
Typical traps. Treating chunk size as one-way (bigger is not always better), and confusing metrics that only check whether a relevant item appears with metrics that also reward good ordering.
What to practice. Chunk the same document three ways and compare what a search returns. Learn what recall, precision, and ranking-aware metrics each tell you.
Domain 3: Application Development (30%)
What it covers. The widest domain. Choosing LangChain-style tools; assessing responses for quality and safety; selecting chunking from evaluation results; augmenting prompts with context from the user’s input; changing a prompt to move a response from baseline to target; guardrails; selecting an LLM or an embedding model (including its context length) from attributes, model cards, and experiment metrics; using MLflow and the Agent Framework for agentic systems; comparing the evaluation and monitoring phases; and multi-agent systems that use Genie Spaces or a conversational API.
In practice. Matching a model to a constraint: a short chunk limit points to a smaller embedding model, and a latency-first design rules out heavy extras.
Typical traps. Picking a model by name recognition, not by the stated constraint, and mixing up evaluation (before release) with monitoring (after release).
What to practice. Read a model card and write down the task category, context length, and size. Run an MLflow experiment and compare two models on a metric.
Domain 4: Assembling and Deploying Applications (22%)
What it covers. Coding a simple chain, and a chain as a pyfunc model with pre- and post-processing; the elements of a RAG application (model flavor, embedding model, retriever, dependencies, input example, signature); registering to Unity Catalog with MLflow; controlling access from serving endpoints; creating and querying Vector Search indexes and configuring them for embeddings count, update frequency, latency, and cost; serving with Foundation Model APIs; batch inference with `ai_query()`; persistent stores for memory; CI/CD (index updates, prompt promotion, component tests); MCP servers; prompt version control; and user-facing interfaces such as Apps, Slack, and Teams.
In practice. A nightly table refresh means the index should update on a schedule and not continuously; a user-facing app must not expose long-lived tokens.
Typical traps. Forgetting that each choice has a cost or latency trade-off, and treating “it works” as the same as “it is promoted safely”.
What to practice. Create a Delta Sync index, sync it, and query it. Log a small chain with a signature and input example, and register it.
Domain 5: Governance (8%)
What it covers. Masking as a guardrail, guardrails against malicious input, using legal and licensing requirements for data sources to avoid legal risk, and recommending an alternative when a data source contains problematic text.
In practice. Remove personal data before text is embedded, and check whether you are allowed to use a document at all.
Typical traps. Applying a control at the wrong stage, for example filtering output when the data should never have entered the index.
What to practice. For one example application, list where data enters, where it is stored, and where it leaves, then name a control for each point.
Domain 6: Evaluation and Monitoring (12%)
What it covers. Choosing an LLM by size and architecture from quantitative metrics; picking metrics to monitor for a deployment; evaluating agents with MLflow scoring and tracing; inference logging; controlling LLM costs; inference tables and Agent Monitoring for live endpoints; judges that need ground truth; AI Gateway for tracking usage; custom scorers; and incorporating expert feedback.
In practice. Offline, you have a curated dataset with expected answers. Live, you usually do not.
Typical traps. Using a judge that needs ground truth on live traffic, and mixing up logging, tracing, and monitoring features.
What to practice. Run an evaluation with two scorers, then list which of them would still work with no expected answers.
More guides for DATABRICKS-GENAI-ENGINEER-ASSOCIATE
- Databricks GenAI Associate Exam Format
- Is the Databricks GenAI Associate Hard?
- Databricks GenAI Associate Prerequisites
- Databricks GenAI Associate Renewal and Retakes
- Is Databricks GenAI Associate Worth It?
- Databricks GenAI Associate Cheat Sheet
- GenAI vs ML Associate: Which Databricks Exam?
Next step
Take the study plan and map it onto these domains, then test yourself with the sample questions. For full-length sets, see our Databricks GenAI Engineer Associate practice tests.
Frequently asked questions
Which domain should I study first?
Start with Design Applications and Data Preparation, because the later domains build on how retrieval works.
Is code tested?
The questions are multiple choice, but they name functions, parameters, and packages. Reading Python (and some SQL) fluently helps.
Do I need to know every Agent Bricks feature?
You need to know when each type fits. The outline asks you to determine how and when to use them.