Databricks Certified Generative AI Engineer Associate

Last checked October 22, 2026 against the official Databricks exam page.

1. Question 1Application Development

   A vector index was built with one embedding model, and incoming queries are embedded with a different model. What is the most likely result?

   1. ARetrieval quality degrades because query vectors and document vectors are not in a comparable space
   2. BThe index re-embeds every document with the query model on each request, so quality stays the same
   3. CSearch becomes faster because the query model uses fewer dimensions to represent the same text
   4. DNothing changes, because the embedding model affects only how vectors are stored and not how they are compared
   Show answer and explanationHide answer

   Correct answer: A

   Similarity only means something when documents and queries are embedded with the same model, so mixing models returns poor matches. The index does not re-embed on each query. Fewer dimensions do not make mismatched vectors comparable, and the embedding model determines the vectors that are compared.

2. Question 2Assembling and Deploying Apps

   A RAG application must be deployed as a REST endpoint, with a review app that stakeholders can use to test answers. Which Databricks capability fits this need?

   1. AA notebook scheduled to run each hour and write its output to a shared dashboard for stakeholders
   2. BAgent Framework deployment, which creates a Model Serving endpoint and a review app for the agent
   3. CPublishing the notebook as a read-only dashboard and sharing its link with the stakeholder group
   4. DUploading the agent's Python file to the workspace as a text file, so that other users can download it
   Show answer and explanationHide answer

   Correct answer: B

   Agent Framework deployment creates a Model Serving endpoint for the agent and a review app for stakeholder testing, in one managed step. A scheduled notebook produces batch output, not an interactive REST endpoint. A dashboard and a text file do not serve the agent to applications.

3. Question 3Data Preparation

   Long technical PDFs are split into fixed-size chunks for retrieval. Answers often miss information that spans a section break. Which change is most likely to help?

   1. AUse chunks that respect section boundaries, with a small overlap between adjacent chunks
   2. BStore each PDF as a single row and retrieve the entire file for every query, regardless of its length
   3. CRemove headings and tables from the source documents, because they add tokens that the model does not need
   4. DSplit on fixed token counts with no overlap, and keep only the first chunk of each section in the index
   Show answer and explanationHide answer

   Correct answer: A

   Respecting section boundaries keeps related content together, and a small overlap preserves context that crosses a split. Keeping only the first chunk of a section discards the rest of it. Removing headings and tables throws away the structure that helps retrieval, and retrieving whole files for every query exceeds context limits.

4. Question 4Design Applications

   A company wants one assistant that answers questions about HR policies and a second capability that books meeting rooms through an internal API. Which design is most appropriate?

   1. AA model fine-tuned on every past booking record, so that it can answer policy questions and make bookings itself
   2. BOne large prompt that answers policy questions from memory and handles bookings in plain text without any tools
   3. CA single agent that retrieves policy passages with RAG and calls a booking function registered as a tool
   4. DA keyword search index over the policy files, with users booking rooms by sending an email to the facilities team
   Show answer and explanationHide answer

   Correct answer: C

   Decomposing the problem into retrieval for policy answers and a registered tool for bookings lets each capability be tested and governed separately. A prompt with no tools cannot perform the booking action, and fine-tuning on booking records does not give reliable access to live room availability. Keyword search returns documents rather than answers, and email booking is not an application design.

5. Question 5Evaluation and Monitoring

   A team has 500 labelled test questions and wants to measure whether a RAG application's answers are correct and grounded, repeatedly, after each change. Which approach fits best?

   1. AAsk one reviewer to read a few answers at random each week and record an impression of overall quality
   2. BRun Mosaic AI Agent Evaluation on the labelled evaluation set and compare the metrics between versions
   3. CTrack only the average response length, because longer answers are usually more complete and correct
   4. DCount the tokens in each response, which indicates how much information the model has provided to users
   Show answer and explanationHide answer

   Correct answer: B

   Agent Evaluation runs the labelled set through the application and reports metrics such as correctness and groundedness, so versions can be compared repeatedly. Random spot checks are slow and subjective. Length and token counts do not measure correctness.