Databricks Certified Generative AI Engineer Associate practice questions

Question 1

A team is being asked which API contract Foundation Model APIs expose for chat completions, completions, and embeddings on a customer-support chatbot. Per the Databricks Foundation Model APIs overview, which option matches the published API format?

Accepted Answer

All endpoints expose an OpenAI-compatible API format supporting chat completions, completions, and embeddings operations.

Answer

All endpoints expose a proprietary RPC format that is explicitly incompatible with the OpenAI chat completions, completions, or embeddings schema.

Answer

All endpoints expose a JDBC-only interface and explicitly do not support REST-based chat completions, completions, or embeddings.

Answer

All endpoints expose only embeddings on REST and require gRPC for any chat completion or completion operation on Databricks.

Question 2

A platform team is being asked to summarize how MLflow standardizes packaging and deployment for an enterprise rolling out many GenAI applications. Per the Databricks MLflow models documentation, which option matches the published workflow scope?

Accepted Answer

MLflow model management provides a standardized workflow for packaging and deploying GenAI applications.

Answer

MLflow model management on Databricks explicitly does not standardize packaging or deployment and forces each team to invent its own workflow.

Answer

MLflow model management on Databricks only supports classical scikit-learn pipelines and explicitly cannot package GenAI applications for deployment.

Answer

MLflow model management on Databricks is a deprecated UI replaced by raw cloud-storage scripts for every deployment workflow.

Question 3

A platform team is being designed a hybrid search experience for a customer-support RAG that needs both keyword precision and semantic recall. Per the Databricks Vector Search documentation, which option matches the published hybrid-search approach?

Accepted Answer

Hybrid keyword-similarity search combines vector embeddings with BM25 relevance scoring through Reciprocal Rank Fusion.

Answer

Hybrid keyword-similarity search runs two independent queries and always returns only the keyword result, discarding the vector ranking entirely.

Answer

Hybrid keyword-similarity search is implemented by averaging the raw cosine score and a Jaccard token overlap with no use of BM25 or Reciprocal Rank Fusion.

Answer

Hybrid keyword-similarity search requires running a separate full-text engine like Elasticsearch outside Databricks because Vector Search does not combine BM25 with embeddings.

Question 4

A company is designing an AI agent for an insurance back office and the engineer must select which capabilities the agent's tools can provide beyond text generation. Per the Databricks AI agent tools documentation, which TWO options match documented agent tool capabilities?

Accepted Answer

Searching document collections and querying databases to ground the agent's responses in real data

Accepted Answer

Calling external REST APIs or running custom code to perform specialized business logic actions

Answer

Retraining the base foundation model from each conversation so future answers improve automatically

Answer

Provisioning new Model Serving GPU endpoints on demand whenever request latency begins to climb

Question 5

A team is being designed automated alerting for an edtech tutoring RAG when response distribution drifts. Per the Databricks monitor-and-diagnose documentation, which option matches the published role of Lakehouse Monitoring on inference tables?

Accepted Answer

Databricks Lakehouse Monitoring creates automated data quality profiles on inference tables, detecting distribution shifts in input patterns and response characteristics.

Answer

Databricks Lakehouse Monitoring explicitly does not analyze inference tables and cannot detect any distribution shift in input patterns or response characteristics.

Answer

Databricks Lakehouse Monitoring requires the customer to write a custom Spark Streaming job per inference table because automated profiling is not provided.

Answer

Databricks Lakehouse Monitoring monitors only structured Delta tables that have nothing to do with AI inference and cannot be applied to a serving inference table.

Question 6

An engineer is preparing to drop a column from a heavily used Unity Catalog table and first needs to identify every downstream table, job, and dashboard that depends on it. Per the Databricks data lineage documentation, which capability supports this requirement?

Accepted Answer

Unity Catalog data lineage, captured automatically to the column level, enabling impact analysis of downstream assets.

Answer

Inference tables, which log Model Serving requests so you can trace which dashboards consumed the column.

Answer

The audit log system table, which records every column read so dependencies can be rebuilt from history.

Answer

Delta Lake time travel, which lets you query an earlier table version to see which columns were referenced.

Question 7

A startup is being designed a cost-conscious AI agent endpoint that is idle most of the day. Per the Databricks Model Serving documentation, which option matches the published cost-control capability for idle endpoints?

Accepted Answer

Endpoints support serverless compute with scale-to-zero capability, eliminating costs during idle periods.

Answer

Endpoints on Databricks always run a minimum of 10 GPUs per region regardless of traffic and explicitly cannot scale to zero during idle periods.

Answer

Endpoints on Databricks bill for compute only when scale-to-zero is disabled, and the scale-to-zero feature does not eliminate costs during idle periods.

Answer

Endpoints on Databricks must be manually paused by an operator every night because the platform has no automated scale-to-zero capability for serving endpoints.

Question 8

An MLOps lead is being asked how to roll out a new RAG model version with automatic rollback if quality regresses. Per the Databricks model serving deployment-patterns documentation, which option matches the published blue-green guidance?

Accepted Answer

Blue-green deployments use traffic routing to gradually shift traffic from old to new model versions, with automatic rollback if error rates exceed configured thresholds.

Answer

Blue-green deployments on Databricks always perform a 100% atomic cutover and explicitly do not support gradual traffic shifting or automatic rollback.

Answer

Blue-green deployments on Databricks require the customer to manually run a kubectl rollout because traffic routing and rollback are not part of Model Serving.

Answer

Blue-green deployments on Databricks support gradual shifting but explicitly do not include any automatic rollback when error rates exceed configured thresholds.

Question 9

An MLOps team is being designed an autoscaling Vector Search endpoint for a customer-support RAG that sees variable diurnal traffic. Per the Databricks Vector Search index documentation, which option matches the published endpoint capability?

Accepted Answer

Vector Search endpoints provide dedicated compute for serving vector search queries with auto-scaling based on query load.

Answer

Vector Search endpoints are a single fixed-size container per region with no autoscaling and no dedicated compute for query traffic.

Answer

Vector Search endpoints share compute with the workspace SQL warehouse pool and explicitly cannot scale independently for diurnal vector query traffic.

Answer

Vector Search endpoints run only on the driver node of the all-purpose cluster that originally created the index and have no separate compute plane on Databricks.

Question 10

A team is designing an AI agent for an edtech tutoring product that must call multiple tools to grade short answers and look up curriculum context. Per the Databricks AI agent framework documentation, which option matches the published agent capability?

Accepted Answer

Agents can execute multi-step reasoning by chaining LLM calls with tool invocations (database queries, API calls, calculations) to accomplish complex tasks.

Answer

Agents on Databricks can only emit a single, deterministic LLM response per user request and explicitly cannot chain multiple tool calls or LLM steps in one task.

Answer

Agents on Databricks must run entirely client-side in the browser, with no server-side LLM call or tool invocation permitted by the agent framework.

Answer

Agents on Databricks compile each step into a Spark job and require a dedicated Photon cluster per agent before any reasoning chain can be invoked.

Question 11

An MLOps team is being asked which dimensions a regulated healthcare RAG must be scored on by the Databricks agent evaluation framework. Per the Databricks agent evaluation documentation, which option matches the published scoring dimensions?

Accepted Answer

Evaluation runs score responses across multiple dimensions: answer correctness verifies factual accuracy against ground truth, faithfulness checks that responses are grounded in retrieved context without hallucination, relevance assesses whether retrieved documents match the query intent, and safety validates content appropriateness.

Answer

Evaluation runs on Databricks score only the response token length and explicitly do not assess correctness, faithfulness, relevance, or safety.

Answer

Evaluation runs on Databricks rely on exact-string match against ground truth and explicitly do not include LLM-as-judge faithfulness, relevance, or safety dimensions.

Answer

Evaluation runs on Databricks score only response latency and explicitly do not evaluate any quality dimension of the response.

Question 12

A team needs to let a data analyst read from a Delta table at sales.transactions.orders in Unity Catalog, but the analyst currently holds only SELECT on the table itself and queries still fail. Per the Databricks Unity Catalog privileges reference, which additional grant resolves the access gap?

Accepted Answer

Grant USE CATALOG on the sales catalog and USE SCHEMA on the transactions schema, plus the existing SELECT.

Answer

Grant the analyst ownership of the sales catalog so that every child privilege cascades down automatically.

Answer

Grant USE CATALOG on the sales catalog only, because schema usage is inherited from the parent catalog grant.

Answer

Grant MODIFY on the orders table, since USE CATALOG and USE SCHEMA are implied by the SELECT grant.

Question 13

An engineer is designing an AI agent whose tools mostly require per-user authentication and the lowest possible execution latency. Per the Databricks AI agent tools documentation, which implementation does Databricks recommend for these most-other use cases instead of Unity Catalog functions?

Accepted Answer

MCP servers, or defining the tool logic directly in the agent code for faster execution

Answer

Wrap each tool as a task-specific AI Function and invoke it from a scheduled SQL batch job

Answer

Register each tool as a separate Mosaic AI Model Serving endpoint and call it over REST

Answer

Unity Catalog functions configured with a JSON schema passed in the tools parameter

Question 14

A platform team is being asked why the AI Playground is unavailable in one of their Databricks workspaces while it works in another region. Per the Databricks AI Playground documentation, which requirement explains the difference?

Accepted Answer

The workspace must be in a region that supports Foundation Models for the AI Playground to be available.

Answer

The workspace must already run a provisioned throughput endpoint before the AI Playground will load for users.

Answer

The workspace must have a dedicated service principal assigned as the AI Playground's recorded creator identity.

Answer

The workspace must have Unity Catalog lineage enabled so the AI Playground can trace model training data.

Question 15

A platform team is designing a Vector Search workflow where the application itself writes and updates embedding vectors directly through code rather than syncing from a Delta table. Which TWO statements correctly describe the Direct Vector Access Index they would use?

Accepted Answer

The user is responsible for updating the index using the REST API or the Python SDK rather than relying on automatic sync.

Accepted Answer

It cannot be created using the UI; you must use the REST API or the SDK to create and maintain this kind of vector index.

Answer

It automatically and incrementally syncs with a source Delta table as the underlying data in that source table actually changes over time.

Answer

It is the only index type that is able to serve hybrid keyword-similarity search results back to the calling application.

Databricks Certified Generative AI Engineer Associate

Sample questions

Sources

Similar exams