1.1 Exam Format, Domains & Scoring

How the Databricks GenAI Associate Exam Is Built

The Databricks Certified Generative AI Engineer Associate exam validates that you can design, build, deploy, govern, evaluate, and monitor large language model (LLM) applications on the Databricks platform. It is deliberately scenario-based: rather than asking you to recite a definition, most items describe a realistic engineering situation — a retrieval chatbot that hallucinates, a slow serving endpoint, a document pipeline full of noisy PDFs — and ask you to choose the best next action. Passing rewards practical judgment far more than memorization, which is why Databricks recommends real hands-on experience before you sit.

The Numbers You Must Know

A small number of unscored (pilot) questions may appear alongside the 45 scored items. They are not flagged, so you should treat every question as if it counts. Because the benchmark is an unrounded 70.00%, the arithmetic is unforgiving but clear: you can miss 13 questions and still pass, but the 14th miss fails you (31/45 = 68.9%). There is no separate per-domain cut score — you need 32 correct answers in total, wherever they come from.

The Six Weighted Domains

The blueprint splits the exam into six domains, and their weights should directly drive how you allocate study hours:

Application Development (30%) and Assembling & Deploying Apps (22%) together are 52% of the exam — about 24 of 45 questions. This is the single most important strategic fact in the entire guide. Retrieval-augmented generation (RAG) chains, agents, tool calling, packaging models with MLflow, configuring Vector Search, and deploying to Model Serving deserve the bulk of your time. Governance is only 8% (roughly four questions), so learn it thoroughly but do not over-invest relative to its weight.

Question Style and Common Traps

Questions come in two forms. Multiple-choice items have one correct answer among four options. Multiple-selection items say something like “Select all that apply” or “Choose two” and require you to pick every correct option with no partial credit — miss one and the whole item is wrong, so these deserve extra care. Distractors are engineered to be plausible-but-suboptimal: a valid approach that is slower, more expensive, less governable, or overkill for the stated requirement. The recurring decision the exam tests over and over is prompting vs. RAG vs. fine-tuning vs. an agent — pick the lightest pattern that satisfies the business need. Use RAG for fresh, proprietary, or citable knowledge; fine-tuning for durable behavior or style; a tool-calling agent only when the app must decide to act on external systems.

Prerequisites and Recommended Experience

There is no formal prerequisite and no education requirement. Databricks recommends related Academy training plus about six months of hands-on experience building GenAI solutions on the platform. All machine-learning code context on the exam is Python; some non-ML data-manipulation context may appear in SQL. The strongest single preparation step is to build one small RAG app end to end so chunking, embeddings, Vector Search, prompting, MLflow packaging, and evaluation become procedural rather than theoretical.

A Worked Time Budget

Ninety minutes for 45 questions is exactly two minutes per question. A reliable plan is a three-pass strategy. Pass one (about 60 minutes): answer every question you know within roughly 90 seconds and flag anything that is a long scenario, a multiple-selection item, or a genuine toss-up. Pass two (about 20 minutes): return to the flagged items, which now feel easier because your mind has been primed by related questions. Pass three (about 8–10 minutes): sanity-check flagged multiple-selection answers and make sure no question is blank. There is no penalty for guessing, so an eliminated-and-committed guess always beats an empty answer.

Common Scoring Mistakes to Avoid

Three traps sink otherwise-prepared candidates. First, treating multiple-selection like multiple-choice — selecting only the single best option when the item needs two or three; there is no partial credit, so a half-right answer scores zero. Second, over-optimizing for the ideal engineering answer when the question stem fixes a constraint (budget, latency, existing tooling) that makes a simpler option correct. Third, spending five minutes on one hard item and running out of time for four easy ones at the end — protect your pace, because every question is worth exactly the same.

Recertification and Validity

The credential is valid for two years. Databricks refreshes exams as the platform evolves, so recertification means passing the then-current version of the exam before your badge expires — there is no shortcut renewal. Because the platform changes quickly, plan to re-engage with new features (like the March 2026 additions) as part of staying certified.

The March 18, 2026 Blueprint Update

Databricks published an interim guide stating the blueprint changes on March 18, 2026. The updated objectives add deeper coverage of Agent Bricks, managed or external MCP (Model Context Protocol) server integration, AI Gateway usage tracking, custom scorers, interactive agent interfaces, and CI/CD for prompts and vector indexes. The six domain names and weights stay the same; the added emphasis lands inside Application Development, Assembling & Deploying, and Evaluation & Monitoring. Confirm your test date: on or after March 18, 2026, study the delta topics as well as the core six domains.