Databricks Generative AI Engineer Associate Practice Questions

Scenario · Select one answer

A warehouse classifier must return exactly `AVAILABLE` when stock is greater than zero and `UNAVAILABLE` otherwise. Which instruction best defines the contract?

1. Return `AVAILABLE` whenever the message mentions a product.

2. Return one of the two labels and include a JSON object with the warehouse and reasoning.

3. Return `UNAVAILABLE` whenever the message contains a quantity.

4. Given the validated stock quantity, return only `AVAILABLE` if it is greater than zero; otherwise return only `UNAVAILABLE`.

Scenario · Select one answer

An internal compliance assistant must answer from policies that are revised several times per month. The team wants the smallest maintainable architecture and does not want a training cycle for every revision. Which component set fits?

1. A retriever, a prompt that inserts retrieved evidence, and a response-generating LLM

2. A scheduled fine-tuning job that retrains the LLM after every policy revision

3. A response LLM with prompt instructions but no access to the policy corpus

4. A multi-agent supervisor and a fine-tuned model even though the task has one fixed retrieval path

Concept · Select one answer

A reporting service must reduce a multi-paragraph incident narrative to one sentence while preserving the principal event and impact. Which NLP task should guide model selection?

1. Text-to-text translation

2. Sentence boundary detection

3. Topic classification

4. Summarization

Scenario · Select one answer

Within the March 2026 exam-guide taxonomy: An enterprise wants one conversational entry point that delegates finance questions to a Genie Agent, policy questions to a Knowledge Assistant, and actions to governed tools. Which Agent Bricks capability fits?

1. Information Extraction

2. Supervisor Agent

3. A single PromptTemplate

4. A Delta Sync index

Code / API · Select 2 answers

A router uses a typed output: class Route(BaseModel): category: Literal["billing", "technical", "cancellation"] route_llm = llm.with_structured_output(Route) Which TWO practices make routing more reliable? (Choose two.)

1. Add every company document to the routing prompt

2. Keep a closed Literal set for the permitted categories

3. Validate the parsed category before invoking the selected branch

4. Ask for free-form reasoning and parse the last word with split()

5. Increase temperature so the model can invent new route names

Scenario · Select one answer

An AI Search query retrieves 20 candidates, but offline labels show that the most relevant passages often appear below the three positions sent to the LLM. What should the team test first?

1. Apply a reranker to the candidate set before selecting the final three passages.

2. Ask the response LLM to reorder passages after the prompt has already discarded seventeen candidates.

3. Replace the embedding model without first measuring a ranking baseline.

4. Send every candidate to the generator regardless of context budget.

Scenario · Select 2 answers

A corpus produces more chunk rows than the selected index can hold. Which TWO configuration changes reduce the row count without removing source documents? (Choose two.)

1. Increase the maximum chunk size within the embedding model's supported limit.

2. Reduce overlap between adjacent chunks while preserving enough boundary context.

3. Decrease chunk size.

4. Increase overlap.

5. Use embeddings with fewer dimensions but keep identical chunk boundaries.

Scenario · Select one answer

A procurement assistant answers invoice-policy questions but cannot report the current delivery milestone for a particular purchase order. Every request includes `purchase_order_id`. Which source closes the gap?

1. A vector index containing generic carrier marketing pages.

2. A governed operational table keyed by `purchase_order_id` with current milestone and estimated arrival.

3. A fine-tuning set of last year's delivery examples.

4. A prompt rule that adds a fixed number of days to every order date.

Code / API · Select one answer

A RAG corpus contains PDFs with headings, tables, lists, and reading order that must be preserved for downstream chunking. Which Databricks function is designed to produce structured document content?

1. ai_query

2. ai_parse_document

3. vector_search

4. ai_classify

Concept · Select one answer

A corpus must answer both broad questions about an entire topic and precise questions about individual passages. Which advanced retrieval preparation is designed for multiple abstraction levels?

1. RAPTOR-style hierarchical summaries

2. Removing all metadata

3. One sentence per index

4. Increasing model temperature

Scenario · Select one answer

An ingestion benchmark counts tokens with each candidate's own tokenizer, including required prefixes and special tokens. Document inputs need at most 512 tokens and separately embedded queries at most 90. All candidates meet the quality threshold. Which candidate should be tested first for minimum footprint without truncation?

1. A 768-dimensional model with a 514-token limit and 0.44 GB footprint.

2. A 2,560-dimensional model with a 2,048-token limit and 11 GB footprint.

3. A 4,096-dimensional model with a 32,768-token limit and 14 GB footprint.

4. A 384-dimensional model with a 512-token limit and 0.13 GB footprint.

Concept · Select 2 answers

A chat assistant must retrieve governed reference material and preserve explicit conversation state across turns. Which TWO capabilities directly satisfy those requirements? (Choose two.)

1. A searchable knowledge store or retriever for the reference material

2. A conversation-state or checkpoint store scoped to the session

3. An unrestricted action tool even though no action is required

4. A document loader used only at runtime in place of an index

5. A front-end component library

Scenario · Select one answer

A commercial application must process newly published Spanish medical documents with an open model from a model hub. Which evidence should drive the shortlist before task-specific evaluation?

1. Membership in a popular model family, regardless of its card.

2. Download count alone.

3. Upload date alone.

4. A model card documenting the task, Spanish and medical-domain evidence, context limits, license, risks, and intended use.

Concept · Select one answer

A RAG response states facts that are not supported by any retrieved document. Which quality issue should the engineer investigate first?

1. Groundedness

2. Throughput

3. Embedding dimension

4. Token billing

Scenario · Select one answer

A team wants to compare paragraph-based chunks with fixed 500-token chunks. What is the soundest experiment?

1. Change chunking, embedding model, and LLM at the same time

2. Use the same labeled queries and retrieval metrics while changing only the chunking strategy

3. Choose whichever produces fewer rows without evaluating answers

4. Ask one user which option looks better

Scenario · Select one answer

A billing tool requires `account_id` and `transaction_id` for every lookup. How should an agent incorporate those values?

1. Fine-tune the model on historical identifiers so it predicts both fields from the user's wording.

2. Extract both fields, validate their format and authorization, and pass them explicitly in the governed tool call.

3. Store the latest identifiers in serving environment variables and reuse them across every user's session.

4. Retrieve semantically similar billing records first and let the generator infer any missing identifiers.

Scenario · Select one answer

A model must emit exactly {"priority": "high|medium|low"}. Zero-shot instructions still produce extra prose. What is the best next prompt change?

1. Add few-shot examples with only valid JSON outputs

2. Increase temperature

3. Ask for chain-of-thought

4. Remove the schema

Code / API · Select one answer

A custom agent is built with a third-party framework but must integrate with Databricks tracing, evaluation, deployment, and monitoring. Which MLflow interface is the recommended wrapper?

1. ResponsesAgent

2. pandas.DataFrame

3. PromptTemplate only

4. AutoMLClassifier

Concept · Select one answer

The same groundedness scorer runs on a fixed regression dataset before release and on sampled live traces afterward. What distinguishes these two uses?

1. Offline evaluation compares versions on controlled examples; monitoring checks changing live traffic over time.

2. Using the same scorer makes both runs offline evaluations, regardless of where their examples originate.

3. Monitoring measures infrastructure only, so a groundedness scorer cannot be used on production traces.

4. Running evaluation before release removes the need to assess answer quality after deployment.

Code / API · Select one answer

A custom agent must call a Genie Agent and maintain a multi-turn conversation with it. Which integration is appropriate?

1. The Genie conversation API

2. Copy every table into the system prompt

3. OCR the SQL warehouse

4. A BLEU scorer

Code / API · Select 2 answers

A public chatbot currently returns the raw model response: response = chain.invoke(request) return response Which TWO changes implement an output safety guardrail? (Choose two.)

1. Store the unsafe response in the system prompt for the next user

2. Evaluate response with a toxicity or safety classifier before returning it

3. Block, redact, or replace responses that exceed the configured risk threshold

4. Increase temperature when toxic text is detected

5. Apply safety checks only to the user's input

Code / API · Select one answer

A current LCEL chain is incomplete: prompt = PromptTemplate.from_template("Explain {topic} in one sentence") chain = prompt | ______ | StrOutputParser() Which replacement completes the runnable prompt → model → parser sequence?

1. prompt.format(topic='RAG')

2. StrOutputParser()

3. AISearchClient()

4. ChatDatabricks(endpoint='support-llm')

Code / API · Select one answer

A notebook uses the current SDK: from databricks.ai_search.client import AISearchClient client = AISearchClient() The source is a Delta table with a text column, and Databricks must compute embeddings and synchronize the index. Which method creates the correct index type?

1. client.get_index(...)

2. client.create_delta_sync_index(...)

3. client.create_direct_access_index(...)

4. client.similarity_search(...)

Scenario · Select one answer

A catalog fits both tested AI Search endpoint types. Under the expected concurrency, measured (p95 latency, retrieval quality) is: standard ANN (70 ms, 0.91), standard with reranking (160 ms, 0.96), storage-optimized ANN (260 ms, 0.92). Requirements: quality ≥0.90 and p95 <100 ms. Which deployment fits the evidence?

1. Standard with reranking, because the highest quality score overrides the latency requirement.

2. Storage-optimized ANN, because a higher maximum vector capacity always means lower query latency.

3. Standard ANN, since its measured quality and latency satisfy both mandatory thresholds.

4. Delay deployment until embeddings are fine-tuned, although the measured standard configuration passes both thresholds.

Scenario · Select one answer

A release pipeline must promote an approved prompt from staging to production without copying mutable text into application code, while keeping rollback history. Which mechanism fits?

1. Keep one unversioned prompt file and overwrite it during every deployment.

2. Register immutable prompt versions in MLflow and move environment aliases only after evaluation and approval.

3. Store the production prompt only in a CI runner's temporary directory.

4. Overwrite a Delta row in place and delete the prior value.

Concept · Select one answer

The March 2026 exam guide refers to Mosaic AI Vector Search, while current Databricks documentation and the product page use Databricks AI Search. How should an engineer interpret these names?

1. AI Search is a separate replacement product that requires migrating every Vector Search index

2. Mosaic AI Vector Search is only a generic L2 algorithm, while AI Search is the managed service

3. AI Search is the endpoint resource and Vector Search is always a separate index product

4. They are the same Databricks product after a rename; legacy SQL and REST identifiers can still contain vector-search wording

Code / API · Select 2 answers

A Delta table must be classified by an existing custom serving endpoint whose input signature requires a structured field named text of type string: SELECT review, ai_query( endpoint => ?, request => ? ) AS label FROM catalog.schema.reviews; Which TWO expressions correctly fill the placeholders? (Choose two.)

1. 'review-classifier'

2. named_struct('text', review)

3. review

4. current_user()

5. vector_search(review)

Scenario · Select one answer

An agent needs governed retrieval from an existing Databricks AI Search index with minimal operational work. Which MCP integration is preferable?

1. A new custom server that reimplements AI Search

2. A browser extension

3. A prompt-only simulation of search

4. A Databricks-managed AI Search MCP server

Scenario · Select one answer

An authenticated agent backend correctly restricts retrieval for each employee. A new shared answer cache uses only the question text as its key. Two employees have different document permissions. What must change before enabling the cache?

1. Keep the shared key because identical wording always implies identical authorized evidence.

2. Isolate cache entries by verified authorization scope and handle permission changes, or disable shared answer caching.

3. Check permissions only on cache misses; cached answers are safe because an authorized employee generated them.

4. Remove document titles from cached answers while retaining all answer content under the shared key.

Scenario · Select one answer

Offline traces show that irrelevant inflammatory passages in the indexed source repeatedly reach the prompt and trigger offensive answers. What is the most direct upstream remediation?

1. Refresh the same uncurated corpus more frequently.

2. Display a notice explaining that offensive responses are possible.

3. Reduce the number of users while leaving the harmful passages indexed.

4. Review, remove, redact, or quarantine the problematic passages before rebuilding the index, then run regression evaluation.

Concept · Select one answer

What principle applies when a dataset license permits research use only but the intended assistant is a paid commercial product?

1. Use it because public download implies commercial permission

2. Remove the dataset name and continue

3. Fine-tune the model so the original license no longer applies

4. Do not use it commercially unless a compatible license or explicit permission is obtained

Scenario · Select one answer

A search request includes department='finance', supplied by the browser. The verified employee belongs only to support. The index contains both departments. How should the backend construct the authorization filter?

1. Trust the requested department because the employee completed sign-in.

2. Derive the permitted scope from trusted entitlements; reject or intersect incompatible requested filters before retrieval.

3. Retrieve finance documents and instruct the model to reveal only non-sensitive sentences.

4. Accept the filter when the employee writes a convincing business reason in the prompt.

Scenario · Select one answer

A multi-step agent sometimes chooses the wrong tool after retrieval. The engineer needs the ordered calls, arguments, outputs, errors, and timing for a failed request. What should be enabled first?

1. MLflow Tracing for the agent and its components

2. Only an aggregate final-answer score with no execution spans

3. Ad hoc print statements that omit nested framework calls

4. Separate applications for every step before inspecting the current trace

Scenario · Select one answer

Four subject-matter experts apply different interpretations of 'complete answer' and rate the same agent traces inconsistently. What should happen before treating their labels as benchmark truth?

1. Replace all experts immediately with an unrubriced LLM judge.

2. Define observable rubric criteria, calibrate reviewers on shared examples, resolve systematic disagreements, and measure agreement.

3. Average every score without investigating why reviewers disagree.

4. Delete every case on which the experts disagree.

Scenario · Select one answer

A support endpoint has both a latency SLO and a response-quality target. Which production metric set is most useful?

1. Track p50 and p95 latency, errors, throughput, token or cost usage, and sampled quality or user feedback.

2. Track only average latency, without tail latency, errors, request volume, cost, or any response-quality signal.

3. Track training loss and parameter count, even though neither reports serving health or quality on production requests.

4. Track only document file size and ingestion time, regardless of endpoint behavior and generated-answer quality.

Concept · Select one answer

A gateway enforces 60 requests per minute for one caller. The caller stays below that rate but runs continuously for days. What does this control guarantee by itself?

1. The caller's accumulated monthly spend can never exceed a fixed euro budget.

2. Request admission is rate-limited; a per-minute limit alone does not impose a total spending cap.

3. Every admitted response will contain fewer tokens than any response admitted before the limit was set.

4. The usage dashboard automatically shuts the service down at a user-defined monetary threshold.