Databricks Generative AI Engineer practice questions and exam guide

A team is selecting a foundation model for a support assistant. Traffic is bursty, prompts average 3,000 tokens, and the assistant must reason over multi-step troubleshooting trees. Which selection criterion should carry the LEAST weight at design time?

1. The model's published context window relative to prompt plus retrieved context length
2. The model's parameter count as an absolute measure of capability
3. Whether the model is available through pay-per-token Foundation Model APIs
4. Measured quality on a task-representative evaluation set the team builds

Answer: B — The model's parameter count as an absolute measure of capability

B) Correct — parameter count is a poor standalone proxy for capability; models of similar size differ widely by training and instruction tuning, so it should not drive selection. A) Context window is a hard constraint: if prompt plus retrieval exceeds it, the design fails regardless of quality. C) Availability on pay-per-token matters directly because traffic is bursty and provisioned throughput would be idle much of the time. D) Task-representative evaluation is the strongest evidence available and should dominate the decision.

A new GenAI workload's biggest design risk is unclear success criteria. The right first step is:

1. Train a custom model for the new workload
2. Select the largest model currently available
3. Define an eval set and target metrics first
4. Provision additional GPU capacity up front

Answer: C — Define an eval set and target metrics first

C) Correct — an evaluation set with explicit targets for groundedness, correctness, and latency is what every later design choice gets measured against. A) Training before knowing the target is premature. B) Model size chosen without metrics wastes budget. D) Buying infrastructure first inverts the order.

A GenAI proposal lists this success criterion: "users are satisfied with the assistant." Before development starts, the FIRST thing the engineer should do is:

1. Select the foundation model and provision a serving endpoint for prototyping
2. Stand up a feedback widget so satisfaction data accumulates during development
3. Ingest and chunk the full corpus first so that retrieval experiments can begin immediately
4. Define measurable targets and a labeled evaluation set that represents real questions

Answer: D — Define measurable targets and a labeled evaluation set that represents real questions

D) Correct — without measurable targets and a representative evaluation set there is no way to compare design options or know when the project is done. A) Choosing a model first commits to an implementation before the bar it must clear is defined. B) A feedback widget is valuable in production but collects nothing useful before there is a working assistant, and thumbs data alone is too sparse to steer design. C) Chunking before knowing what questions matter means retrieval experiments have nothing to be scored against.

Optimal RAG chunk size range is typically:

1. 10-50 tokens, one short phrase per chunk
2. 5,000+ tokens, several full pages in each chunk
3. 200-1,000 tokens with a small overlap window
4. Exactly one sentence per chunk, no overlap

Answer: C — 200-1,000 tokens with a small overlap window

C) Correct — this range balances retrieval precision against enough surrounding context, and modest overlap preserves continuity at boundaries. A) Too small to carry meaning. B) Too large, so retrieval returns mostly irrelevant text. D) Single sentences strip the context the model needs.

A team reports recall at k=5 of 0.62 and precision at k=5 of 0.55. Which interpretation is MOST defensible?

1. Retrieval is fine; the generator is the problem and should be replaced
2. About a third of questions never see the right passage
3. The evaluation set is too small to be meaningful at these values
4. Precision below recall proves that top-k is set too low

Answer: B — About a third of questions never see the right passage

B) Correct — recall of 0.62 means the supporting passage is absent for about 38 percent of questions, and no generator can answer those correctly, so retrieval is the binding constraint. A) With recall that low the generator is being blamed for missing evidence it never received. C) Set size is a fair general concern but nothing in the numbers indicates it, and the recall gap is large enough to act on. D) The relationship between precision and recall values does not by itself imply a top-k setting; raising k usually raises recall and lowers precision.

For chunking, what does overlap solve?

1. It reduces the total number of tokens embedded
2. It reduces the storage the index consumes
3. It speeds up the embedding computation step
4. It preserves context across chunk boundaries

Answer: D — It preserves context across chunk boundaries

D) Correct — repeating tokens at the seam keeps an entity mentioned near a split retrievable from either neighboring chunk. A) Overlap increases the token count. B) It increases stored data. C) Embedding more text is marginally slower.

An application must call a third-party proprietary model that Databricks does not host, while keeping a single governed interface for the app. The MOST appropriate mechanism is:

1. An external model endpoint in Model Serving, with credentials stored as a secret
2. A Unity Catalog function that issues an HTTP request to the provider from SQL
3. Embedding the provider SDK inside the chain and reading the key from an environment variable
4. A provisioned throughput endpoint configured with the provider's model name

Answer: A — An external model endpoint in Model Serving, with credentials stored as a secret

A) Correct — external model endpoints expose a third-party provider through the same Model Serving interface, so governance, rate limiting, and logging apply uniformly. B) Hand-rolling HTTP calls from a SQL function bypasses the serving layer's routing, credential handling, and usage tracking. C) Embedding the SDK and reading keys from the environment scatters credentials and loses centralized control. D) Provisioned throughput applies to models Databricks hosts; you cannot reserve capacity on a provider Databricks does not serve.

A team already computes embeddings in an external system with a proprietary model and simply needs a serving-grade index to query them. Which Vector Search index type is MOST appropriate?

1. Delta Sync index with a Databricks-managed embedding endpoint
2. Delta Sync index with self-managed embeddings in a source column
3. Direct Vector Access index, with the client writing vectors and metadata
4. A standard Delta table with an approximate nearest neighbour SQL function

Answer: C — Direct Vector Access index, with the client writing vectors and metadata

C) Correct — Direct Vector Access is the mode for full client control of index contents, which suits vectors produced entirely outside Databricks. A) Managed embeddings would require Databricks to compute the vectors, discarding the proprietary model. B) Self-managed embeddings in a Delta Sync index are a reasonable alternative, but they require landing the vectors in a governed source table and accepting sync semantics, which is more coupling than the requirement implies. D) There is no substitute here for a purpose-built vector index; this option invents a workflow rather than using the service.

Why split prompts into system + user + assistant roles?

1. Spark requires these role fields when calling any serving endpoint
2. Roles encrypt the message content in transit to the model provider
3. Roles compress the conversation into fewer billable prompt tokens
4. Chat-tuned models follow roles, separating policy from user input

Answer: D — Chat-tuned models follow roles, separating policy from user input

D) Correct - chat models are trained on the role convention, so system instructions, user input, and prior assistant turns stay distinguishable. A) Spark has nothing to do with the chat message format. B) Roles are plain metadata, not a security control. C) Roles add structure and a few tokens; they do not compress anything.

Which practice does the MOST to improve a Genie space's answer accuracy?

1. raising the SQL warehouse size so that generated queries finish faster
2. curating a small set of well-described tables and adding example question-and-SQL pairs
3. adding every table in the catalog to the space so that no user question is ever out of scope
4. instructing users to phrase every question as valid SQL before submitting

Answer: B — curating a small set of well-described tables and adding example question-and-SQL pairs

B) Correct — Genie quality is driven by metadata quality and worked examples over a tight table scope. A) Warehouse size affects latency, not correctness. C) A wide scope increases ambiguity and join errors — the opposite of what helps. D) Requiring SQL from users removes the reason to use Genie.

Function calling in an agent means the LLM:

1. Returns JSON naming a tool and its arguments
2. Directly executes Python functions in its runtime
3. Runs shell commands on the serving host itself
4. Edits system files to persist its own state

Answer: A — Returns JSON naming a tool and its arguments

A) Correct — the model emits structured JSON identifying a tool and arguments; the orchestrator runs it and feeds the result back. B) The agent code executes, never the model. C) The model has no shell. D) Nothing in function calling touches the host filesystem.

A team wants each sub-agent in a supervisor system to be independently versioned and released. Which arrangement BEST supports this?

1. store each sub-agent's prompt text inside the supervisor's own configuration file
2. bundle all of the sub-agents into a single model artifact and version that bundle
3. keep the sub-agents as notebook functions that the supervisor imports at runtime
4. register and serve each sub-agent as its own Unity Catalog model and endpoint

Answer: D — register and serve each sub-agent as its own Unity Catalog model and endpoint

D) Correct — separate registered models and endpoints give each team its own lifecycle, rollback and monitoring boundary. B) One artifact forces lockstep releases, the opposite of the requirement. C) Runtime notebook imports have no versioning or promotion story. A) Prompt strings in a config file are not a deployable, testable agent boundary.

Provisioned throughput on Model Serving is BEST for:

1. Spiky, low-volume workloads with idle periods
2. High, steady traffic with predictable latency
3. Free experimentation in a development workspace
4. Pre-training a foundation model from scratch

Answer: B — High, steady traffic with predictable latency

B) Correct — reserved capacity pays off when demand is constant, giving stable latency and a predictable bill. A) Sparse traffic wastes reserved capacity; pay-per-token fits better. C) Reserved capacity is billed whether used or not. D) Serving endpoints do not pre-train models.

Which statement about model signatures for agents is MOST accurate?

1. a signature is optional metadata that the serving layer simply ignores at runtime
2. signatures apply only to classical ML models and not to agents
3. the signature declares input and output schema and is validated at request time
4. the signature is inferred fresh on each incoming request payload

Answer: C — the signature declares input and output schema and is validated at request time

C) Correct — the signature is the enforced request/response contract for the served model. A) Serving validates against it rather than ignoring it. B) Agents are logged as models and carry signatures too. D) Inference of the signature happens once, at logging time.

Which concern is MOST specific to exposing an agent through a Teams or Slack surface rather than an internal web app?

1. the agent must be logged to MLflow before it can be invoked from a chat channel
2. the vector search index must be rebuilt separately for every channel used
3. chat-platform identity must be mapped to a workspace identity for data access
4. the prompts must be rewritten in the chat platform's own markup language

Answer: C — chat-platform identity must be mapped to a workspace identity for data access

C) Correct — the identity bridge is the crux: without mapping, everyone effectively shares one service principal's data access. A) MLflow logging is required regardless of the front end. B) Indexes are shared, not per-channel. D) Minor formatting differences are not the substantive concern.

Granting an analyst access to query a model via SQL function should be done by:

1. Sharing the exported model file over cloud storage
2. Granting the analyst cluster administrator rights
3. Sharing the serving endpoint URL and an access token
4. Granting EXECUTE on a UC function wrapping ai_query

Answer: D — Granting EXECUTE on a UC function wrapping ai_query

D) A UC function plus an EXECUTE grant gives the analyst exactly one capability, auditable and revocable, without exposing the endpoint. A) Handing over the artifact bypasses governance entirely. B) Cluster admin is far broader than the task requires. C) A raw URL and token offer no fine-grained control or audit trail.

To demonstrate compliance, the team should retain:

1. Inference Tables, MLflow records, lineage, Gateway logs
2. Free-form engineering notes kept in a shared team drive
3. The Slack channel history where the incidents were raised
4. Verbal confirmation from the engineers who reviewed it

Answer: A — Inference Tables, MLflow records, lineage, Gateway logs

A) These four together cover requests, model provenance, asset relationships, and policy enforcement for the mandated retention period. B) Informal notes are neither complete nor tamper-evident. C) Chat history is unstructured and usually expires. D) Verbal confirmation leaves no record at all.

Unity Catalog governance for AI tools provides:

1. Faster query execution against governed tables
2. Central access control, lineage, and auditing
3. Cheaper compute for jobs that read the tables
4. Higher accuracy for models registered under it

Answer: B — Central access control, lineage, and auditing

B) Correct — Unity Catalog applies one permission, lineage, and audit model across tables, models, vector indexes, and functions. A) Query speed comes from compute and layout. C) Governance does not reduce compute cost. D) Registration does not change model quality.

An endpoint must enforce per-request token limits + provider routing. Right Databricks feature?

1. Delta Live Tables pipelines with expectations on the inputs
2. AI Gateway routes with rate limits and provider fallbacks
3. A Genie space configured with a token budget per user
4. Auto Loader ingesting request logs into a Delta table

Answer: B — AI Gateway routes with rate limits and provider fallbacks

B) AI Gateway is the policy layer in front of endpoints: it defines routes, enforces usage and rate limits, and abstracts the underlying provider. A) DLT expectations validate pipeline data, not request policy. C) Genie does not front model endpoints. D) Auto Loader ingests files after the fact.

A regression slips into production. Best first action:

1. Wait to see whether the metrics recover on their own
2. Disable the application until a new model is trained
3. Roll back to the prior alias and inspect stored traces
4. Re-train the model from scratch on refreshed data

Answer: C — Roll back to the prior alias and inspect stored traces

C) Restore known-good behavior first by repointing the alias, then diagnose offline using the captured request and response traces. A) Waiting extends user impact with no plan. B) A full shutdown is a heavier outage than a rollback. D) Re-training is premature before the cause is known.

When evaluating an agent end-to-end vs component-wise, you should:

1. Only end-to-end scoring, since users see the whole
2. Only component scoring, since it is cheaper to run
3. Both: components localize faults, end-to-end shows UX
4. Neither, relying on user complaints as the signal

Answer: C — Both: components localize faults, end-to-end shows UX

C) Component metrics tell you which stage failed while end-to-end metrics reflect what the user actually experienced, so both are needed. A) End-to-end alone leaves you unable to diagnose. B) Component alone can look healthy while the answer is poor. D) Complaints are a lagging and biased sample.

Which domain carries the LARGEST weight on the current exam guide?

1. Governance — securing and controlling GenAI assets
2. Data Preparation — chunking and embedding source data
3. Application Development — building the agent itself
4. Evaluation and Monitoring — judging and tracking quality

Answer: C — Application Development — building the agent itself

C) Correct — Application Development is the heaviest domain at 30 percent. A) Governance is the smallest at 8 percent. B) Data Preparation is 14 percent. D) Evaluation and Monitoring is 12 percent.

Which statement about the format of the Databricks Certified Generative AI Engineer Associate exam is accurate?

1. it has 45 scored items within a 90-minute window and no lab component
2. it is a four-hour hands-on practical lab exam in a live workspace
3. it is a set of 100 items answered over an unlimited amount of time
4. it requires submitting a portfolio project for review by an expert panel

Answer: A — it has 45 scored items within a 90-minute window and no lab component

A) Correct — 45 scored items in 90 minutes, online proctored, with unscored pretest items possible. B) There is no lab component on this exam. C) Both the item count and the timing are wrong. D) No portfolio submission exists for this certification.

A candidate asks what score they must achieve to pass. The accurate answer is that

1. a fixed 70 percent is required on every Databricks associate exam
2. Databricks does not publish a passing score for this certification
3. a candidate must pass each domain of the exam independently
4. the score is curved against the other candidates taking it that month

Answer: B — Databricks does not publish a passing score for this certification

B) Correct — no cut score is published, so any specific figure circulating is unofficial. A) States an unpublished number as fact. C) There is no per-domain pass requirement; scoring is overall. D) No published curve or cohort-relative scoring exists.

Common RAG failure modes?

Poor retrieval (wrong docs), too few/many chunks, conflicting sources, model ignoring context, hallucination, latency overruns.

What does a Mosaic AI Model Serving endpoint provide for a registered agent?

A scalable REST endpoint with autoscaling and optional scale-to-zero, authentication and access control, request and response logging to inference tables, and traffic splitting across model versions.

What is chunking in a RAG pipeline and why is it necessary?

Chunking splits source documents into retrievable passages sized for the embedding model and the generator's context window. Without it, documents are too large to embed faithfully and retrieval returns too much irrelevant text.

What licensing review is required for source data used in RAG or tuning?

Confirm you hold the right to store, embed and reproduce the content in generated answers; check third-party and scraped content terms, honor attribution or non-commercial restrictions, and exclude data whose license forbids derivative or machine-learning use.

Latency budgets?

Plan: tokens-out × per-token latency + retrieval + tool calls. Stream responses to reduce perceived latency.

What role does a Genie Space play inside a multi-agent system?

It answers natural-language questions over governed structured tables by generating SQL, and can be attached as a tool or sub-agent so the supervising agent can hand off quantitative questions instead of guessing from unstructured text.