Free Databricks GenAI Engineer sample questions
Try these before reading the answers. The mix follows the real weighting, so most come from development and deployment.
Question 1. A RAG application returns fluent answers that cite the wrong section of a long technical manual. Retrieval returns five chunks per query. What is the most likely cause?
- A. The model is too small for technical content
- B. Chunking strategy — chunks are too large or lack overlap, so retrieval matches poorly
- C. The temperature setting is too high
- D. The maximum output length is too short
Show answer
Answer: B
Fluent answers from the wrong source point at retrieval rather than generation, and oversized chunks with no overlap are the usual cause because relevant passages get split or diluted. A larger model would summarise the wrong context just as fluently, lowering temperature does not change what is retrieved, and more output tokens changes length only.
Question 2. An application must answer from a document corpus that is updated daily. The team wants the lowest ongoing effort. Which approach fits?
- A. Fine-tune the model nightly on the updated corpus
- B. Select a model with a larger context window
- C. RAG over an index refreshed as documents change
- D. Paste the full corpus into the system prompt
Show answer
Answer: C
Retrieval-augmented generation reads the current index at query time, so a daily-updated corpus only requires the index to be refreshed rather than any model work. Fine-tuning daily is expensive and repeats forever, a larger context window does not supply the documents, and pasting everything into the prompt does not scale to a corpus.
Question 3. A team needs to track which model version, parameters and metrics produced a deployed generative AI application, and manage its lifecycle. Which Databricks capability is central?
- A. MLflow
- B. Delta Live Tables
- C. A vector search index
- D. Photon
Show answer
Answer: A
MLflow provides experiment tracking, model versioning and lifecycle management, which is what recording parameters and metrics against a deployed version requires. Unity Catalog governs access and lineage, Delta Live Tables builds data pipelines, and a vector index stores embeddings for retrieval.
Question 4. Stakeholders ask for a number showing how often the assistant's answers are actually supported by the documents retrieved. Which evaluation measure reports this?
- A. Average response latency
- B. Token cost per query
- C. Relevance
- D. Groundedness
Show answer
Answer: D
Groundedness measures whether a response is supported by the retrieved source content, which is exactly the claim being made. Latency measures speed, token cost measures spend, and relevance only checks that the answer addresses the question rather than that it is supported.
Question 5. An application must ensure that users can only retrieve content from documents their team is permitted to see, and that access is auditable. Which Databricks capability governs this?
- A. MLflow model registry
- B. Unity Catalog
- C. A model serving endpoint
- D. A carefully written prompt template
Show answer
Answer: B
Unity Catalog provides centralised access control and lineage over data and models, which is what permission-scoped retrieval with an audit trail requires. MLflow tracks experiments and models, a serving endpoint hosts the application, and a prompt template shapes the request without enforcing permissions.
How did you do?
Notice that three of the five are about the application around the model rather than the model itself. That matches the exam: development and deployment are 52% of it, and the questions assume you have built a RAG application rather than read about one.
If question 1 felt uncertain, spend time on chunking — see week 1 of the study plan.