About these questions
These five questions are original. We wrote them to match the style of the Databricks Generative AI Engineer Associate exam outline, which has six domains: Design Applications, Data Preparation, Application Development, Assembling and Deploying Applications, Governance, and Evaluation and Monitoring. These questions cover five of them. They are not real exam questions, and the real exam will word things differently. Platform features change, so always check the Databricks documentation for current behavior.
How to use these
Try each question without looking at your notes, and pick an answer before you open the explanation. Then read the explanation even when you were right, because the reason the closest wrong option fails is where the learning is. Note any domain where you slipped and revisit it on the domains explained page.
Question 1 (Design Applications)
An insurance team is building an assistant over several hundred policy wordings. Their engineers have already prototyped a parent-child splitter that keeps each clause with its section heading. They need to keep that splitter, swap embedding models as they test, and keep orchestration in the framework their other services already use. Which approach fits?
- A. Create an Agent Bricks Knowledge Assistant over a volume holding the policy documents.
- B. Create an Agent Bricks Information Extraction task over the policy documents.
- C. Create a Genie Space over the volume that holds the policy documents.
- D. Build the agent by hand with the Mosaic AI Agent Framework, choosing the splitter, embedding model, and framework, and log it with MLflow.
Show the answer and explanation
Correct answer: D. The team needs control over chunking, the embedding model, and the orchestration framework. A hand-built agent gives that control, and MLflow can log and manage it afterward. A Knowledge Assistant, the closest wrong option, is a quick managed way to answer questions over documents, but it does not let the team plug in its own splitter or framework. Information Extraction pulls structured fields out of documents instead of answering open questions. A Genie Space is aimed at questions over structured data, not a library of policy text.
Question 2 (Data Preparation)
A retrieval team compares two configurations on the same query. Both return exactly the same five documents, but in a different order. Users of the search page read only the first two or three results. Which metric will tell the two configurations apart?
- A. Recall at 5
- B. Precision at 5
- C. Recall at 20
- D. NDCG at 5
Show the answer and explanation
Correct answer: D. NDCG rewards putting the most relevant documents near the top, so two orderings of the same five documents can score differently. Recall at 5 and precision at 5 only look at which documents appear in the top five. Since both configurations return the same set, these two metrics give identical numbers. Recall at 20 is the closest wrong option: a larger cutoff still counts the same documents, so it also cannot separate them. The point to remember is that set-based metrics ignore order, while ranking-aware metrics do not.
Question 3 (Application Development)
A lender’s assistant must end every answer about loan offers with the exact sentence “Rates may change without notice.” and must never use the phrase “risk-free”. The check is purely mechanical. It must run on every row of the offline evaluation set and later on sampled production traces, and the team does not want an extra LLM call per row. What should they build?
- A. A Guidelines judge with both rules written in natural language
- B. A Correctness judge with the required sentence as the expected answer
- C. A Python function decorated with `@scorer` that checks the output text and returns a pass or fail
- D. An ExpectationsGuidelines judge with the rules stored in the dataset
Show the answer and explanation
Correct answer: C. A code-based scorer is plain Python. It checks for the exact sentence and the forbidden phrase deterministically, costs no model call, and can run on both the evaluation set and production traces. The Guidelines judge is the closest wrong option because it states the rules clearly, but it is an LLM judge, so it adds a model call per row and may be less strict than exact string matching. Correctness compares an answer against a reference answer, which is not what this check does. ExpectationsGuidelines is also an LLM judge, so it breaks the no-extra-call requirement as well.
Question 4 (Assembling and Deploying Applications)
A team creates a Delta Sync index and wants Databricks to generate the embeddings from a text column. Which two parameters tell Databricks which text to embed and which model to embed it with? Choose two.
- A. `embedding_vector_column`
- B. `embedding_model_endpoint_name`
- C. `embedding_dimension`
- D. `embedding_source_column`
- E. `pipeline_type`
Show the answer and explanation
Correct answers: B and D. With managed embeddings, `embedding_source_column` names the text column to embed and `embedding_model_endpoint_name` names the embedding model endpoint to use. `embedding_vector_column` and `embedding_dimension` belong to the other style of index, where you compute the vectors yourself and supply them. The closest wrong option is `pipeline_type`, because it is also set when you create a Delta Sync index. It controls whether the index syncs continuously or when triggered, not what gets embedded or with which model.
Question 5 (Evaluation and Monitoring)
A team monitors a customer-support assistant in production. Live traffic carries no expected answers, and scoring every request would cost too much, so they score a sample. Which two built-in judges can they run on this traffic? Choose two.
- A. Correctness
- B. Safety
- C. RetrievalSufficiency
- D. RelevanceToQuery
Show the answer and explanation
Correct answers: B and D. Safety and RelevanceToQuery are reference-free judges. They look at the response and the question, not at an expected answer, so they work on live traces. Correctness and RetrievalSufficiency need ground truth, meaning expected facts or an expected response. Correctness is the closest wrong option: it is the most familiar judge, but without expected answers it has nothing to compare against. Judges that need ground truth belong in offline evaluation on a curated dataset.
More guides for DATABRICKS-GENAI-ENGINEER-ASSOCIATE
- Databricks GenAI Associate Exam Format
- Databricks GenAI Associate Study Plan
- Is the Databricks GenAI Associate Hard?
- Databricks GenAI Associate Prerequisites
- Databricks GenAI Associate Renewal and Retakes
- Is Databricks GenAI Associate Worth It?
- Databricks GenAI Associate Cheat Sheet
- GenAI vs ML Associate: Which Databricks Exam?
Next step
Revisit any domain where you were unsure on the domains explained page. For more questions in this style, with an explanation for every option, try our Databricks GenAI Engineer Associate practice tests.