CertifyAIxAce every AI certification.

DB-GENAI-ASSOC sample questions with answers

10 free practice questions for the Databricks Certified Generative AI Engineer Associate exam. Try each one, then open the answer to see why the right option wins and every other option loses.

Question 1Design Applications

An engineer is writing a nightly job that applies a generative step to every row of a very large table with aiquery. Worried about throughput, they plan to split the table into many small batches by primary-key range and drive one aiquery statement per batch from an orchestrator, and to point the calls at a provisioned throughput endpoint they will create for the job. What does the Databricks guidance actually recommend here?

  1. A.

    Keep the provisioned throughput endpoint but run one statement per batch on a dedicated job cluster sized to the batch count.

  2. B.

    Replace ai_query with individual REST calls to the serving endpoint so that the orchestrator can control the concurrency directly.

  3. C.

    Submit the full dataset in a single query against a Databricks-hosted foundation model, letting AI Functions handle parallelism, retries and scaling.

  4. D.

    Keep the manual batching but move to a Databricks-hosted foundation model, because the split is what provides the parallelism.

Show answer

Answer: C

Submit the whole dataset in one query on a Databricks-hosted foundation model and let AI Functions manage scaling.

  • A. Provisioned throughput is not the recommended target for ai_query batch work, and per-batch statements remain discouraged.
  • B. Hand-built REST concurrency discards the automatic retry and scaling behaviour that AI Functions provide.
  • C. This matches both documented recommendations: one query over the full dataset against a Databricks-hosted foundation model.
  • D. The manual split is the part the guidance argues against; AI Functions already parallelise the work.
Question 2Design Applications

An architect must decide how to expose two capabilities to an agent. The first is a fixed parameterised lookup: given an account number and a date range, return that account's settled transactions. The second is an open-ended need to answer whatever question a user asks about the same warehouse. Which assignment follows the documented guidance?

  1. A.

    A Unity Catalog SQL function for the fixed lookup, and a Genie Agent or the SQL MCP server for the open-ended questions.

  2. B.

    A Unity Catalog SQL function for both, since a single mechanism keeps governance and discovery consistent.

  3. C.

    A Genie Agent for both, since it can be constrained with instructions to behave like a fixed lookup when needed.

  4. D.

    A custom MCP server for the fixed lookup, and a Unity Catalog SQL function for the open-ended questions.

Show answer

Answer: A

Unity Catalog functions suit queries known in advance; open-ended questions belong to Genie or the SQL MCP server.

  • A. This matches the documented split between predetermined parameterised queries and open-ended natural-language questions.
  • B. A fixed function cannot serve questions whose SQL is not known when the function is written.
  • C. Routing a settled, deterministic lookup through natural-language generation loses determinism for no benefit.
  • D. The assignment is inverted and adds a custom server for work Unity Catalog already governs.
Question 3Design Applications

An agent on a Databricks-hosted Claude endpoint must first call a pricing tool and then return a strictly schema-conformant object describing the quote. The developer tries to send tools and a jsonschema responseformat together in one request so that a single round trip does both, and the request is refused. Which design keeps both the tool call and the guaranteed output shape?

  1. A.

    Send one request with the tools attached to obtain the tool result, then send a second request carrying the response_format and no tools to render the final object.

  2. B.

    Drop response_format and restate the schema inside the system prompt, since a schema written in prose is equivalent once the model has the tool result.

  3. C.

    Declare the output object as an additional tool and rely on the model choosing to call it last, which yields validated arguments.

  4. D.

    Keep both parameters and set tool_choice to none so the endpoint stops treating the request as a tool-calling request.

Show answer

Answer: A

Claude on Databricks cannot combine response_format with tools, so the tool turn and the formatting turn must be separate calls.

  • A. Separating the tool turn from the formatting turn satisfies the documented restriction while keeping schema enforcement.
  • B. A schema written in prose is an instruction, not a constraint, and reintroduces the malformed-output failure mode.
  • C. A pseudo-tool for output relies on the model deciding to call it, so the shape is still not enforced by the endpoint.
  • D. tool_choice is itself named in the restriction, so setting it to none does not make the combination legal.
Question 4Design Applications

A finance operations team has millions of archived supplier invoices as PDFs and needs a typed table of supplier, invoice number, dates and totals produced once, with no conversational interface and no user waiting on a response. A colleague suggests standing up an Agent Bricks offering because the work is information extraction. What should the team build instead?

  1. A.

    A Supervisor Agent whose single subagent is an AI Search index built over the scanned invoice images.

  2. B.

    A batch SQL pipeline using aiparsedocument to recover document structure and ai_extract with a typed schema.

  3. C.

    A custom agent with a code interpreter tool that opens each PDF and reads the fields programmatically.

  4. D.

    A Knowledge Assistant over the invoice volume, queried programmatically once per invoice through its agent endpoint.

Show answer

Answer: B

Bulk document extraction with no conversation is AI Functions work: parse the documents, then extract against a schema.

  • A. A supervisor with one subagent orchestrates nothing, and indexing for search is not extraction.
  • B. Parsing with ai_parse_document then extracting with a typed ai_extract schema is the documented document-processing path.
  • C. Hand-rolling PDF reading in a code tool rebuilds the layout parser the platform already provides.
  • D. Calling a chatbot endpoint once per invoice uses a conversational surface as a batch API and loses batch scaling.
Question 5Design Applications

A Supervisor Agent works perfectly for its author, who can reach every underlying asset. After it is shared with a wider audience, some users find that questions answerable only by one particular subagent come back incomplete, while everything else behaves normally. Nothing in the supervisor's configuration differs between the two groups. What should the team check first?

  1. A.

    Whether each end user has been granted explicit access to every subagent, which the documentation states is required.

  2. B.

    Whether the affected users are in a region where the supervisor's model serving endpoint is unavailable.

  3. C.

    Whether the supervisor exceeds the documented cap on how many subagents one system may contain.

  4. D.

    Whether the subagent's description is too similar to another subagent's, causing the supervisor to misroute.

Show answer

Answer: A

End users need explicit access to each subagent; the supervisor does not pass through the author's privileges.

  • A. The documentation requires end users to have explicit access to each subagent, which explains the per-user difference.
  • B. A regional serving restriction would break the supervisor as a whole, not one topic for some users.
  • C. Exceeding the subagent cap would affect every user identically rather than only those lacking a grant.
  • D. Ambiguous descriptions cause misrouting for everyone, including the author, which the stem excludes.
Question 6Design Applications

A deployed agent must receive, alongside every user message, the caller's business unit and preferred units of measure, so that retrieval filters and answer formatting can respect them. A developer proposes prefixing each user message with a line such as "BU=EMEA; units=metric" and parsing it back out inside the agent. Which approach does the Databricks agent authoring guidance support instead?

  1. A.

    Store the values in a Unity Catalog table keyed by user and have the agent look them up on every turn.

  2. B.

    Put them in the system prompt at logging time, since they are configuration rather than per-request data.

  3. C.

    Encode the values in the endpoint name so that each business unit queries a separately deployed copy of the agent.

  4. D.

    Pass them as custom inputs on the request, which the agent reads from the request's custom inputs field.

Show answer

Answer: D

ResponsesAgent supports custom inputs on the request, which is where per-call context belongs.

  • A. A lookup adds a round trip and cannot express a preference the caller chooses on a particular request.
  • B. Values fixed in the system prompt at logging time are the same for every caller and cannot vary per request.
  • C. One endpoint per business unit multiplies deployments and versions for what is a single request parameter.
  • D. ResponsesAgent documents custom inputs on the request as the place for context beyond the chat messages.
Question 7Design Applications

Field engineers query a parts knowledge base with a mixture of natural-language symptoms and literal part codes such as XK-4471B. Pure semantic search retrieves plausible articles about similar symptoms but frequently misses the article that contains the exact code, because the code embeds into nothing meaningful. Which retrieval configuration in Databricks AI Search addresses both halves of the query?

  1. A.

    Switch the index to a larger embedding model so that alphanumeric codes gain distinguishable representations.

  2. B.

    Increase the number of results returned by the approximate nearest neighbour query and rely on the reranking stage.

  3. C.

    Apply a metadata filter on the part code column so that only articles carrying that code are considered.

  4. D.

    Use hybrid keyword-similarity search, which combines vector and keyword matching and fuses the two result lists.

Show answer

Answer: D

Hybrid keyword-similarity search combines vector and keyword retrieval, which is what a mixed symptom-and-code query needs.

  • A. No embedding model reliably distinguishes arbitrary part codes, and reindexing to find out is costly.
  • B. A longer candidate list adds no lexical matching, so the exact-code article can still be missing entirely.
  • C. A filter assumes the code has already been parsed into a structured column and ignores the symptom half of the query.
  • D. Hybrid keyword-similarity search combines vector and keyword retrieval and fuses the ranked lists.
Question 8Design Applications

A Knowledge Assistant answers well for most of a department's corpus, but several documents are never cited and the assistant behaves as though it has never seen them. Those files sit in the same Unity Catalog volume as the rest, are in a supported format, and open correctly. An engineer notices that the affected filenames all begin with an underscore. What is the most likely explanation?

  1. A.

    The assistant deduplicates by content hash, and those documents are near-duplicates of files it already indexed.

  2. B.

    Unity Catalog treats underscore-prefixed files as internal metadata and hides them from every reader.

  3. C.

    Files whose names begin with an underscore or a period are skipped during ingestion, as the documented limitations state.

  4. D.

    An underscore prefix causes the files to sort ahead of the rest, and only the first page of the listing is indexed.

Show answer

Answer: C

Knowledge Assistant's documented limitations include skipping files whose names start with an underscore or a period.

  • A. Content-hash deduplication is not a documented behaviour, and nothing suggests these files duplicate others.
  • B. Unity Catalog does not hide such files from readers generally; the skipping belongs to this ingestion path.
  • C. The documented limitations state that files starting with an underscore or a period are skipped during ingestion.
  • D. No pagination limit on the file listing is documented, so sort position does not determine what is indexed.
Question 9Design Applications

A team is assembling a SQL-first retrieval pipeline over a corpus of policy documents held in a Unity Catalog volume. They want to keep as much of the work as possible inside AI Functions rather than writing bespoke Python, and they are mapping each stage of the pipeline onto the function that performs it. Which two mappings match the documented purpose of the functions named? (Select two.)

Choose 2.

  1. A.

    Turning the parsed documents into search-ready chunks for a retrieval pipeline is the job of aiprepsearch.

  2. B.

    Assigning an embedding model to the index is the job of ai_enrich, which infers it from the schema.

  3. C.

    Producing a per-chunk confidence score for retrieval quality is the job of aianalyzesentiment.

  4. D.

    Creating and refreshing the underlying Databricks AI Search index is the job of ai_similarity.

  5. E.

    Retrieving ranked documents and grounded answers from a knowledge source is the job of ai_search.

Show answer

Answer: A, E

aiprepsearch chunks documents for retrieval; ai_search returns ranked documents and grounded answers.

  • A. ai_prep_search is documented as converting documents into search-ready chunks for retrieval pipelines.
  • B. ai_enrich generates columns from a schema and does not configure an index's embedding model.
  • C. ai_analyze_sentiment measures the tone of text and reports nothing about retrieval quality.
  • D. ai_similarity compares two strings; index creation and refresh are Databricks AI Search operations.
  • E. ai_search is documented as retrieving ranked documents and grounded answers from knowledge sources.
Question 10Design Applications

A platform team has standardised on Unity Catalog functions for every agent tool because of catalog governance and discoverability. Reviewing the current documentation, an architect notes that Databricks now steers most tools elsewhere. Which three properties does the documentation attribute to MCP servers and agent code that motivate that recommendation?

  1. A.

    Lower inference cost, automatic retry on tool failure, and schema validation of every tool response.

  2. B.

    Offline availability, cross-workspace replication, and automatic version pinning of each registered tool.

  3. C.

    Guaranteed tool ordering, deterministic argument binding, and built-in caching of repeated tool results.

  4. D.

    Faster execution, support for per-user authentication so a tool can act as the person asking, and additional flexibility compared with Unity Catalog function tools.

Show answer

Answer: D

The documented advantages are faster execution, per-user authentication support and greater flexibility.

  • A. Retry behaviour and response schema validation are not presented as distinguishing properties of MCP servers.
  • B. Offline availability and cross-workspace replication are not documented characteristics of this choice.
  • C. Tool ordering follows from the model's reasoning and is not guaranteed by the transport carrying the tool.
  • D. These are the three advantages the documentation names when steering most tools away from catalog functions.

Keep going with 594 more DB-GENAI-ASSOC questions

Free papers every day, in the real exam formats, with progress by exam domain. Unlock every paper and timed mock exam when you are ready.

Practise freeSee all DB-GENAI-ASSOC papers
DB-GENAI-ASSOC sample questions with answers (10 free) · CertifyAIx