DGAEA
Free DGAEA practice questions
10 questions from our DGAEA practice tests, each with the correct answer and a written explanation. Read them with the answers shown, or switch to quiz mode and answer first.
Question 1 of 10, Design Applications
An HR department wants a chatbot that answers employee questions from about 2,000 policy PDFs stored in a Unity Catalog volume, with citations, and wants it running with the least custom code. Which Databricks option fits BEST?
- AA Genie space built on a Delta table listing the policy PDF file names
- BAgent Bricks Knowledge Assistant pointed at the volume containing the policy documents Correct answer
- CAgent Bricks Information Extraction configured to pull policy fields into a structured table
- DA custom LangGraph agent with a hand-built chunking job, Vector Search index and serving endpoint
Answer and explanation: Knowledge Assistant is the Agent Bricks offering for question answering with citations over unstructured documents, and it handles parsing, indexing and serving with minimal code. Information Extraction turns documents into structured fields rather than answering open questions. Genie answers questions over structured tables, and a table of file names contains none of the policy content. A custom LangGraph RAG agent would work but requires far more custom code than the requirement allows.
Question 2 of 10, Design Applications
A team is designing an agent that answers questions about a customer's open orders. The agent must look up the customer, retrieve their orders and then check shipment status with a carrier API. How should the tools be ordered?
- ACarrier status call first so delivery data is cached, then the customer lookup, then the order query that filters on shipped items
- BOrder query first across all customers, then the customer lookup to filter the result set, then the carrier status call for the matches
- CAll three tools in parallel with the raw user question so the agent can merge whichever results come back first
- DCustomer lookup first, then the order query that needs the customer ID, then the carrier status call that needs the tracking numbers Correct answer
Answer and explanation: Each tool depends on an output of the previous one: orders need the customer ID, and the carrier API needs tracking numbers from the orders, so the sequence must follow that data dependency. Calling the carrier first is impossible because no tracking numbers exist yet. Querying all orders before identifying the customer is wasteful and risks exposing other customers' data. Running the tools in parallel ignores the dependencies, so the later calls would lack required inputs.
Question 3 of 10, Data Preparation
A team indexes long technical manuals organised into chapters and numbered sections. Retrieved chunks often begin mid-sentence and mix the end of one section with the start of the next. Which chunking change BEST addresses this?
- ASplit on section headings first, sub-splitting only sections that exceed the token limit Correct answer
- BKeep the current fixed-size splitter but raise the overlap between adjacent chunks to 50 percent
- CIncrease the fixed chunk size to 4,000 tokens so that most sections fit entirely within one chunk
- DSwitch to fixed 128-token chunks with no overlap for uniform size
Answer and explanation: Structure-aware chunking that follows section headings keeps each chunk semantically coherent and avoids boundaries in the middle of a section, with a fallback split only for oversized sections. Smaller fixed chunks would cut sentences even more often. Very large fixed chunks still ignore section boundaries and dilute the embedding with unrelated text. Heavy overlap duplicates content and inflates the index without aligning chunks to the document structure.
Question 4 of 10, Application Development
Which statement correctly distinguishes evaluation from monitoring in the lifecycle of a generative AI application?
- AEvaluation tracks request volume and latency in production, while monitoring compares candidate models on a fixed test set
- BEvaluation and monitoring are the same activity and differ only in whether the results are stored in MLflow or Delta
- CEvaluation happens only once at project start, while monitoring is used only to retrain the base model on new user data
- DEvaluation scores quality on curated datasets before release; monitoring tracks live traffic Correct answer
Answer and explanation: Evaluation is the offline, pre-deployment assessment against evaluation datasets, and monitoring continuously observes production behaviour such as quality, latency and cost. The second option reverses the two phases. Evaluation is repeated iteratively during development, and monitoring is not limited to retraining. They are distinct phases with different data sources and purposes.
Question 5 of 10, Application Development
An assistant must answer questions about the user's own calendar. The user's ID and time zone are known from the session, but users write questions like 'Am I free tomorrow afternoon?' What should the application add to the prompt so the model can answer correctly?
- AThe complete calendar of every employee so the model can compare free time across people
- BThe user's time zone and today's date, plus calendar events fetched for that user ID Correct answer
- CA list of all public holidays in every country so the model can rule out unusual dates
- DA reminder that the model should ask the user to restate the question with exact timestamps
Answer and explanation: Resolving 'tomorrow afternoon' requires the current date and the user's time zone, and answering requires that user's events, so all three should augment the prompt. Including every employee's calendar leaks data and adds noise. A global holiday list is irrelevant to the user's availability. Forcing users to restate questions with timestamps defeats the purpose of the assistant.
Question 6 of 10, Application Development
A multi-agent system must let a supervisor agent answer questions such as 'What were Q3 sales by region?' using governed sales tables, while another sub-agent handles document questions. What is the MOST appropriate way to give the supervisor access to the sales data?
- AExpose a Genie space over the sales tables as a tool via the Conversation API or managed MCP Correct answer
- BPaste the full contents of the sales tables into the supervisor's system prompt so it can calculate the totals itself
- CFine-tune the supervisor's LLM on the sales tables every night so that it memorises the latest figures by region
- DExport the sales tables to CSV files, chunk them, and add them to the same Vector Search index used for the documents
Answer and explanation: Genie translates natural-language questions into SQL over governed tables, and exposing it through the Conversation API or managed MCP lets a supervisor delegate structured-data questions. Chunking tables into a vector index loses the ability to aggregate accurately. Pasting tables into a prompt does not scale and breaks with context limits. Nightly fine-tuning cannot reliably memorise numbers and is costly and stale.
Question 7 of 10, Assembling and Deploying Applications
A custom model endpoint created without declaring resources fails with a permission error when its chain calls a Vector Search index. Re-logging the model is not possible today. What is the MOST appropriate interim fix?
- AAsk end users to send their own tokens in each request body so that the chain can authenticate with them
- BEmbed the engineer's personal access token directly in the chain code and redeploy the same model version
- CUse a service principal whose secret is referenced in an environment variable, granted SELECT on the index Correct answer
- DGrant SELECT on the index to all users in the account so that the endpoint no longer needs any specific identity
Answer and explanation: Using a dedicated service principal whose credentials are stored as a secret and exposed through an environment variable, with least-privilege SELECT on the index, restores access securely until the model can be re-logged with resources. Granting the index to everyone violates least privilege and the endpoint still lacks credentials. Personal tokens in code are insecure and tied to one person. Passing user tokens in request bodies exposes credentials.
Question 8 of 10, Assembling and Deploying Applications. Choose 2.
Which TWO statements about Mosaic AI Vector Search are correct? (Choose two.)
- ADirect Vector Access indexes are refreshed automatically from a source Delta table on a schedule
- BVector Search supports only exact nearest-neighbour search and cannot use approximate methods
- CA Delta Sync index can compute embeddings for a text column using a model serving endpoint Correct answer
- DEvery index must be stored outside Unity Catalog in a separate workspace-level vector store
- EA vector search endpoint hosts one or more indexes and is the compute that serves queries Correct answer
Answer and explanation: Endpoints are the serving compute that host indexes, and a Delta Sync index with managed embeddings computes embeddings from a source text column by calling an embedding endpoint. Indexes are Unity Catalog objects, not separate workspace stores. Direct Vector Access indexes are written through the API and have no source table to sync from. Vector Search uses approximate nearest-neighbour algorithms for scale.
Question 9 of 10, Governance
An internal assistant should help employees but must refuse requests to generate malware or instructions for violence, regardless of which LLM the agent calls. Which control BEST enforces this?
- AGive the endpoint a strict rate limit of five requests per minute per user
- BScreen inputs and outputs with a safety classifier such as a Llama Guard model inside the agent Correct answer
- CAppend a polite refusal sentence to every document stored in the Vector Search retrieval index
- DGrant CAN QUERY on the endpoint to members of the security team only
Answer and explanation: A dedicated safety classifier applied to both inputs and outputs detects harmful requests and responses independently of the generation model, so the policy holds whichever LLM is used. Appending text to documents does not reliably make the model refuse. Rate limits control volume, not harmful content. Restricting access to the security team would block the employees the assistant is meant to serve.
Question 10 of 10, Evaluation and Monitoring
An agent is evaluated with MLflow. The team has questions but no reference answers yet. Which built-in judge can they run WITHOUT ground truth?
- ARetrievalSufficiency, which checks that context contains the expected facts
- BRelevanceToQuery, which checks whether the response addresses the user's request Correct answer
- CCorrectness, which compares the response with expected facts written by experts
- DA judge that scores semantic similarity to a labelled expected response
Answer and explanation: RelevanceToQuery judges the response against the request alone, so it needs no labels. Correctness compares against expected facts or an expected response, which requires ground truth. RetrievalSufficiency checks whether retrieved context covers the expected facts, which also requires labels. Similarity to an expected response obviously needs a labelled reference.