DB-GENAI-ASSOC study guide: Databricks GenAI Engineer

The Databricks Certified Generative AI Engineer Associate exam tests whether you can design, build, deploy and govern LLM applications, RAG pipelines and agents on the Databricks platform. It is an associate-level certification for engineers with hands-on Databricks experience. You answer 45 scored multiple-choice questions in 90 minutes, online or at a test centre.

DB-GENAI-ASSOC at a glance

Details as of September 2026 — confirm on the official exam page before booking.

Who this exam is for

The exam is for people who build LLM-enabled solutions on Databricks: breaking a problem into components, choosing models and tools, and using the platform's vector search, model serving, MLflow and Unity Catalog to ship RAG applications and agents. There is no prerequisite, but Databricks highly recommends related training and six or more months of hands-on experience with the tasks in the exam guide.

Databricks does not publish a passing score, because it sets scores statistically and adjusts them as questions change. The certification is valid for two years, and recertifying means passing the current version of the exam.

What the exam covers

The current exam guide is dated March 2026. Databricks renames products often, and the guide already lags the documentation in places. Vector Search is now documented as Databricks AI Search, AI Gateway as Unity Gateway, Genie Spaces as Genie Agents, and Agent Framework as Custom Agents. Learn the operations and concepts; they carry across renames.

Design Applications (14%)

This domain is about turning a business goal into an AI pipeline: specifying inputs and outputs, selecting the model task, choosing chain components, and ordering tools for multi-step reasoning. It includes designing a prompt that returns a specific format, such as valid JSON, and deciding when an Agent Bricks option (Knowledge Assistant or a supervisor agent) solves the problem without custom code.

Data Preparation (14%)

Everything that feeds retrieval: chunking strategies matched to document structure and model limits, stripping content that degrades RAG quality, picking a Python package to extract text from a given source format, and writing chunked text into Delta tables in Unity Catalog in the right order. It also covers choosing source documents, measuring retrieval quality, advanced chunking and the role of re-ranking.

Application Development (30%)

The heaviest domain. You select LangChain or similar tools, write and adjust prompts, augment prompts with context from the user's input, implement guardrails, and choose models: by application attributes, by model card, by experiment metrics, and by embedding context length. It also covers building agentic systems with MLflow, letting multi-agent systems query data through Genie, and the difference between evaluation before release and monitoring after it.

Assembling and Deploying Applications (22%)

The most code-shaped domain. Expect to read chains built as pyfunc models with pre- and post-processing, choose the pieces of a RAG application (model flavour, embedding model, retriever, dependencies, signature, input example), register models to Unity Catalog with MLflow, and create and query a vector search index configured for update frequency, latency and cost. It also covers serving with Foundation Model APIs, batch inference with ai_query(), access control from serving endpoints, agent memory, MCP servers, prompt versioning, CI/CD and user interfaces such as Apps, Slack or Teams.

Governance (8%)

A short domain on masking as a guardrail, defending against malicious user input, respecting licensing terms of data sources, and handling problematic text in source data.

Evaluation and Monitoring (12%)

You choose models from quantitative metrics, select monitoring metrics for a deployment, and evaluate agents with MLflow scoring and tracing. Know which evaluation judges need ground truth, how inference tables and usage tables track a live endpoint, how to control LLM costs, how custom scorers work, and how subject-matter expert feedback improves an agent.

A 4-week study plan

This plan assumes six to eight hours a week in a Databricks workspace. Application Development and deployment together make up just over half the exam, so they get two weeks.

Week 1: Design and data preparation (14% + 14%). Build a small RAG corpus. Try two chunking strategies on the same documents and compare retrieval results. Write the chunks into a Delta table in Unity Catalog. Do mixed papers on both domains.

Week 2: Application Development (30%). Build a chain with a prompt template, a retriever and a guardrail. Change the prompt to force a structured output. Compare two models on the same evaluation set in MLflow and note which metrics decided it.

Week 3: Assembling and deploying (22%). Log the chain as an MLflow model with a signature and input example, register it to Unity Catalog, serve it, and query it. Create a vector search index and practise choosing between sync options. Run a batch job with ai_query().

Week 4: Governance, evaluation and rehearsal (8% + 12%). Review guardrails, masking and licensing, then evaluation judges, inference tables and cost controls. Sit a full-length mock in 90 minutes, drill weak topics, then sit a second. For review technique, see [how to use practice exams effectively](/blog/how-to-use-practice-exams-effectively).

Where candidates lose marks

• Learning the guide's product names only. The documentation now uses newer names for several products. If you have only seen the old names, a correct answer can look unfamiliar. Learn what each component does.
• Getting the RAG build order wrong. Questions often ask for the sequence: extract, clean, chunk, write the chunks to a Delta table in Unity Catalog, then build and sync the index that embeds them for retrieval. One step out of place makes the whole option wrong.
• Misreading code. Pyfunc and index-query snippets hide the answer in a detail, such as a missing signature, the wrong model flavour, or pre-processing applied at the wrong point. Read the snippet line by line before the options.
• Confusing evaluation with monitoring. Offline evaluation with judges and ground truth happens before release; inference tables and usage tracking watch a live endpoint. Task 3.12 tests exactly this boundary.
• Choosing a bigger model by reflex. Model-selection questions usually reward the smallest model that meets the quality metric, because cost and latency are part of the requirement.

Practising for DB-GENAI-ASSOC on CertifyAIx

CertifyAIx has 604 original questions for this exam, written against the March 2026 guide and weighted by its six domains, with explanations for every option. Read [how CertifyAIx practice questions are written](/blog/how-certifyaix-practice-questions-are-written) for the method.

• 3 full-length mock exams, timed at the real exam's 90 minutes
• 27 mixed domain papers and 29 topic papers of up to 25 questions each (60 minutes per 25)
• Code-reading questions for chains, indexes and serving

Start with the [DB-GENAI-ASSOC practice questions](/certifications/databricks-generative-ai-engineer-associate).

Frequently asked questions

How much experience do I need?

Databricks recommends six or more months of hands-on experience with the tasks in the exam guide, and related training. There is no formal prerequisite.

How long should I prepare?

If you already build on Databricks, three to five weeks is a sensible range. If you are new to the platform, allow longer, because the deployment domain assumes you have registered, served and queried models yourself.

Are CertifyAIx questions real Databricks exam questions?

No. Every question is original, written from Databricks' public exam guide. Using leaked exam content risks your certification; see [why exam dumps put your certification at risk](/blog/why-exam-dumps-put-your-certification-at-risk).

What is the passing score?

Databricks does not publish it. It sets passing scores through statistical analysis and says they can change as the exam is updated.

How long does the certification last?

Two years. To recertify, you take the current version of the exam.

CertifyAIx is independent and not affiliated with Databricks. Databricks Certified Generative AI Engineer Associate is a trademark of its owner. All CertifyAIx practice questions are original.