Databricks GenAI Engineer study plan

Five weeks at 6–8 hours a week, assuming Databricks recommends 6+ months of hands-on generative AI experience and you have roughly that. You need a Databricks workspace you can build in — this exam rewards having shipped something.

The plan follows the weighting. Application development and deployment are 52% between them; governance is 8%.

Week 1: design applications (14%) and data preparation (14%)

The two smaller front-end sections together.

Design:

• Framing a generative AI problem: what the application must do, for whom, with what constraints.
• Choosing between prompting, retrieval and fine-tuning for a stated requirement.
• Selecting a model against capability, cost and latency.
• Designing the components of a RAG application end to end.

Data preparation:

• Chunking — size and overlap, and why both affect retrieval quality directly.
• Embeddings and vector representation of content.
• Extracting and cleaning source documents.
• Building and maintaining a vector index.

Chunking is worth real attention. It is the most consequential and most overlooked decision in a RAG application, and it appears in both this section and application development.

Weeks 2 and 3: application development (30%)

The largest section. Two weeks.

Week 2 — the core pattern:

• Building a RAG application on Databricks, start to finish.
• Retrieval: how queries match content, and why answers go wrong when retrieval does.
• Prompt construction, including system instructions and how retrieved context is inserted.
• Handling context window limits.

Week 3 — chains and tools:

• LLM chains — multi-step pipelines where output feeds the next step.
• Tool and function calling.
• Agents, and the controls that belong around actions.
• Handling failure: what happens when retrieval returns nothing, or a tool errors.

Week 4: assembling and deploying apps (22%)

• Packaging an application for deployment on Databricks.
• Serving endpoints and what they cost.
• MLflow for tracking and managing what you deploy.
• Unity Catalog for governing models and data.
• Scaling, latency and cost considerations in production.

Week 5: governance (8%), evaluation and monitoring (12%), then practice

The two smallest sections plus revision.

Governance:

• Access control over models and data through Unity Catalog.
• Guardrails on what the application will and will not do.
• Handling sensitive data in prompts and retrieved content.
• Lineage and auditability.

Evaluation and monitoring:

• Measuring quality when there is no single right answer.
• Groundedness, relevance and correctness against retrieved content.
• Building an evaluation set and comparing versions.
• Monitoring in production: quality, latency, cost, and drift in the corpus.

Then: a full practice exam under real conditions, mistakes sorted by section, weakest rebuilt, and a second practice exam.

Where the hours go

Build one RAG application

If you do one thing, do this. Take a folder of documents, chunk and embed them, build an index, construct a chain that retrieves and answers, deploy it to an endpoint, register it in Unity Catalog, and run an evaluation over it.

That single project touches every section on the exam. It is also the most efficient preparation available, because Databricks writes questions from the perspective of someone who has done exactly that.

Deliberately break your retrieval — oversized chunks, no overlap — and watch answers stay fluent while becoming wrong. That failure mode is the most examined idea in the largest section.