Databricks GenAI Engineer resources

Application development carries 30% of this exam and assembling and deploying another 22%. Over half the marks therefore reward having built and shipped a retrieval application rather than read about one. Databricks Academy covers the syllabus free; the build is what makes it stick.

Start with the official material

The exam guide and its weights: application development 30%, assembling and deploying apps 22%, design applications 14%, data preparation 14%, evaluation and monitoring 12%, governance 8%.

Databricks Academy’s generative AI material, which is free self-paced and written to the exam.

The vector search and model serving documentation, since the deployment half of this exam turns on specifics that summaries flatten.

Build a RAG application end to end

This is the highest-value preparation by a distance.

• Chunk a real document set badly on purpose — whole documents as single chunks — and observe retrieval returning topically related but useless passages. Then fix it with smaller overlapping chunks. That failure mode is an exam favourite.
• Preserve table structure through preparation and see what is lost when you do not.
• Implement hybrid search and find a query where keyword matching catches an identifier that semantic search missed.
• Constrain output to a schema and validate it, rather than asking the model politely for JSON.
• Attach a tool that does something consequential and require confirmation before it executes.
• Feed in a document containing a prompt injection and watch it work, then isolate retrieved content from the instruction channel and watch it stop.
• Deploy with traffic split between two prompt versions and compare measured outcomes.
• Register the full configuration — prompt, model version, retrieval settings — so the behaviour is reproducible six months later.

The passing score problem

Databricks does not publish passing scores. Any figure quoted elsewhere is a guess. A weighted practice test is your only calibration — the free 20-question test here matches the published sections.

Worth paying for

A practice test with explanations.

Compute for a real corpus. Retrieval problems do not appear on ten documents. They appear on ten thousand, and this exam asks about them.

What to skip

Braindumps.

Model training theory. This is an application engineering exam, not a model development one — that is ML Associate and ML Professional.

Generic prompt engineering content. The exam tests retrieval architecture, deployment and evaluation far more than prompt wording.

A sensible order

1. Read the exam guide and note that development plus deployment is 52%.
2. Work through the Academy generative AI material.
3. Build the RAG application, including the deliberate failures above.
4. Take the practice test.
5. Revisit the weakest section in documentation.

Where this sits

The Databricks certification path places this alongside the ML lane rather than above it — it is an associate exam in its own right, not a follow-on from ML Associate.