Databricks GenAI Associate Cheat Sheet

The short version

This sheet is a last-pass review in our own words. It does not cover everything on the exam, and it is not a replacement for the exam guide or the Databricks documentation. Use it to spot gaps, then go back to the full domains explained page. Facts are from the exam guide dated March 18, 2026, checked October 22, 2026.

Exam facts at a glance

Item	Detail
Scored questions	45, multiple choice and multiple selection
Unscored items	Up to 10, unmarked, time already added
Time	90 minutes
Passing score	Not published
Aids	None
Languages	English, Japanese, Brazilian Portuguese, Korean
Validity	2 years, renew by taking the current exam
Retake	After 14 days

Domain weights: Design 14%, Data Preparation 14%, Application Development 30%, Assembling and Deploying 22%, Governance 8%, Evaluation and Monitoring 12%.

Domain 1: Design Applications

Start from the business goal. Write down the input, the output, and the steps between them before you pick tools.
A prompt that must return a fixed shape should say the shape (for example, named JSON fields) and give an example.
Agent Bricks types, in one line each: Knowledge Assistant answers questions over your documents; Information Extraction pulls structured fields out of text; Multiagent Supervisor routes work between specialist agents.
Decision rule: if a managed option meets the need, it usually beats a hand-built one on effort.
Trap: choosing the most powerful design when the scenario asks for the simplest that works.

Domain 2: Data Preparation

Chunk size and overlap: smaller chunks give more records and precise matches but less context; larger chunks give fewer records and more context but may exceed a model’s limit. More overlap keeps context across boundaries but creates more records.
Match the extraction tool to the file type: scanned images need OCR; web pages need tag stripping.
Remove boilerplate (menus, footers, repeated headers) before embedding, because it dilutes retrieval.
Chunks are written to Delta tables in Unity Catalog, so access control follows the table.
Recall-style metrics ask whether the right item was retrieved; ranking-aware metrics also reward putting it near the top. Re-ranking reorders retrieved results with a second, usually slower, model.
Trap: assuming bigger chunks are always better.

Domain 3: Application Development

Pick an embedding model whose context length fits your chunks. A chunk longer than the limit is cut off.
Pick a model from its card and from metrics you measured, not from its name.
Evaluation happens before release on a prepared dataset; monitoring happens after release on live traffic.
Guardrails are checks around the model. An input guardrail screens what goes in; an output guardrail screens what comes out.
Decision rule: if the scenario puts latency or cost first, prefer the smaller model or the lighter step.
Trap: improving the prompt when the real problem is missing data in the index.

Domain 4: Assembling and Deploying Applications

A chain can be logged as a pyfunc model with pre- and post-processing. Log it with an input example and a signature, then register it in Unity Catalog.
Delta Sync index: stays in step with a source Delta table. With managed embeddings Databricks computes the vectors from a text column; with self-managed embeddings you supply a vector column. Direct Vector Access index: you write vectors yourself and there is no source table.
Triggered sync updates the index when you start it and costs less; continuous sync keeps it fresh within seconds and costs more. Standard endpoints need change data feed on the source table.
`ai_query()` runs a model over many rows in SQL, so think batch inference when a scenario mentions a table of rows.
Promote prompts across environments with version control, as you would code.
Trap: choosing continuous sync for a table that refreshes once a night.

Domain 5: Governance

Masking hides sensitive values before they reach the model or the index.
Control the point where data enters. Filtering output is too late if the data should never have been indexed.
Check the license and legal terms of a data source before you use it.
Trap: applying the right control at the wrong stage.

Domain 6: Evaluation and Monitoring

MLflow tracing records what happened inside a run; scorers and judges grade it.
Judges that need no reference answer include Safety and RelevanceToQuery. Judges that compare to expected facts, such as Correctness, need ground truth.
RetrievalRelevance asks whether retrieved documents fit the request; RetrievalSufficiency asks whether they hold what is needed to answer, and needs expectations.
Live traffic usually has no expected answers, so use reference-free judges there.
Inference tables and AI Gateway record usage and help with cost and rate limits.
Trap: choosing a ground-truth judge to score live traffic.

Confusable pairs

Pair	Quick difference
Evaluation / monitoring	Before release on prepared data / after release on live data
Triggered / continuous sync	Run on demand / always on
Managed / self-managed embeddings	Databricks computes vectors / you supply them
Input / output guardrail	Screens the request / screens the response
Retrieval relevance / sufficiency	Are documents on topic / do they contain the answer

Last-day checklist

Reread the exam outline once and mark any bullet you cannot explain.
Redo the questions you missed in practice sets, and read the explanations.
Say aloud how a RAG chain runs from question to answer.
Check the exam guide for changes (Databricks asks you to do this about two weeks ahead).
Check your booking time, system test and ID. See the exam day guide.
Sleep. Do not start new material.

More guides for DATABRICKS-GENAI-ENGINEER-ASSOCIATE

Next step

Test the weak spots with the sample questions, then see the exam format page for timing. For longer practice, try our Databricks GenAI Engineer Associate practice tests. The exam overview links all pages.

Try these practice exams