Free Databricks Certified Generative AI Engineer Associate practice test: 30 questions with answers

Similarity only means something when documents and queries are embedded with the same model, so mixing models returns poor matches. The index does not re-embed on each query. Fewer dimensions do not make mismatched vectors comparable, and the embedding model determines the vectors that are compared.

Few-shot examples plus an explicit schema and a rule against extra text show the model the exact output shape, which makes parsing reliable. Higher temperature increases variation, which works against parsing. Background and tone instructions do not constrain the format, and removing instructions leaves the format to chance.

Registering the Unity Catalog function as a tool gives the agent a typed interface, and governance applies to the function itself. Pasting code into prompts is fragile and exposes logic. Estimating status is a guess, and embedding a personal token in a page exposes credentials.

A rolling summary of older turns keeps the important facts while recent turns stay exact, which bounds the context size and cost. Sending the whole transcript keeps growing the cost. Resetting removes memory entirely, and a larger output limit does not change what the model receives as input.

Hybrid search adds keyword scoring, which matches exact tokens such as model codes, to the semantic signal. A larger dense model still struggles with exact codes. Summaries lose the codes themselves, and random sampling does not retrieve the matching product.

Agent Framework deployment creates a Model Serving endpoint for the agent and a review app for stakeholder testing, in one managed step. A scheduled notebook produces batch output, not an interactive REST endpoint. A dashboard and a text file do not serve the agent to applications.

Autoscaling with scale-to-zero stops replicas when idle, which removes compute cost during quiet periods and scales up for bursts. A fixed fleet and a continuously running cluster both pay for idle time. Manually disabling the endpoint breaks availability and does not adapt to demand.

Databricks Asset Bundles describe resources as code and deploy the same definition to multiple targets, which keeps environments consistent. Manual exports and copies depend on people and drift over time. An archive emailed to an administrator is not a repeatable deployment path.

Registering each version gives a named, immutable record, and pointing the endpoint at a specific version makes rollback a single change. Overwriting in place or storing text in a comment loses the history needed for rollback. Deploying the latest notebook revision makes each release depend on whatever was saved last.

A Databricks secret scope stores the key outside code, and the deployment references it so the value is not written in source or outputs. Notebook constants and READMEs expose the key to anyone who reads the code. Cluster environment variables are readable by users of the cluster, so they are not a private store.

Respecting section boundaries keeps related content together, and a small overlap preserves context that crosses a split. Keeping only the first chunk of a section discards the rest of it. Removing headings and tables throws away the structure that helps retrieval, and retrieving whole files for every query exceeds context limits.

A Delta Sync index keeps the index in step with a Delta source table, and Change Data Feed carries the changes so the sync can be incremental. A Direct Vector Access index is updated by the application rather than synced from a table. File listing with cron is fragile, and endpoint autoscaling changes compute, not index content.

Filtering on a metadata column applied in retrieval means documents from other departments are never returned to the model, which is enforcement rather than a request. A prompt instruction can be bypassed by a crafted question. One index per user does not scale, and removing names from text does not stop retrieval from matching content.

Deduplicating by content hash at ingestion removes exact copies before they reach the index, which fixes the cause. Retrieving more chunks spreads the duplicates across the results rather than removing them. Changing embedding size or generation model does not change the stored duplicates.

Removing boilerplate stops navigation text from competing with real content during retrieval, and metadata lets queries be filtered to the right source or version. File names alone lose most of the content. One chunk per document makes very long chunks that dilute relevance, and a thousand results overload the model's context.

Decomposing the problem into retrieval for policy answers and a registered tool for bookings lets each capability be tested and governed separately. A prompt with no tools cannot perform the booking action, and fine-tuning on booking records does not give reliable access to live room availability. Keyword search returns documents rather than answers, and email booking is not an application design.

A smaller model with a constrained output format and batched requests meets high volume and low latency at lower cost for a fixed label set. The largest reasoning model is slow and expensive at this volume, and training from scratch is far beyond what the task needs. Web-searching agents add latency and do not fit a closed set of labels.

Groundedness checks whether each claim in the answer is supported by the retrieved context, which is the failure described. Perplexity measures how well the model predicts text in general and does not check the source. Latency and token counts are operational measures that say nothing about unsupported claims.

Fine-tuning helps when a consistent format or style is hard to hold through prompts alone, and when the base model handles the domain poorly and labelled examples exist. User volume, branding and region do not change whether the model needs to learn new behaviour. Retrieval handles knowledge that changes, so those factors do not favour training.

Splitting the work into policy retrieval, a live balance tool and an escalation tool keeps each responsibility testable and lets live data come from the source of truth. Pasting balances into a prompt goes stale within hours, and a fine-tuned model cannot know current balances. A static FAQ page offers no automated handling at all.

Agent Evaluation runs the labelled set through the application and reports metrics such as correctness and groundedness, so versions can be compared repeatedly. Random spot checks are slow and subjective. Length and token counts do not measure correctness.

Lakehouse Monitoring computes profile and drift metrics over the inference table, so changes in request data and output quality are tracked over time. Cluster autoscaling and Spark stage timings describe compute behaviour rather than data drift. Notebook history records activity and not model inputs.

Falling groundedness and falling retrieval relevance on production traffic are direct quality signals that a change or drift has harmed answers. Training-cluster CPU, a renamed notebook and a new project member do not show anything about answer quality.

Running both prompts on the same labelled set with the same metrics gives a fair, repeatable comparison. Hand-picked questions and length are biased or irrelevant to correctness. Judgement by the authors of the change is not an independent evaluation.

Per-step latency in traces shows where time is spent in a multi-step agent, which points to the step that slowed down. Notebook counts, repository size and wiki pages describe the team's work rather than request time. They have no bearing on why answers are slow.

Unity Catalog lineage records how tables, models and functions are connected, which supports tracing from source data to the deployed application. Cluster logs record compute events, not data flow. Job and notebook histories show runs and edits, not the relationships between governed assets.

Column-level controls and masking remove personal data before the model can see it, which enforces the rule at the data layer. A prompt instruction can be bypassed by a crafted question. Encrypting answers does not stop disclosure to people who have the password, and a rate limit does not stop a single disclosure.

Separate grants for query and change, given to different groups, match least privilege and keep the write path with the people who maintain the data. Catalog ownership gives far more than needed. Naming conventions do not enforce anything, and sharing a personal login removes accountability.

Unity Catalog audit logs and system tables record access to governed objects, which is the evidence an auditor needs. Driver logs record Spark execution rather than data access by user. Notebook comments and dashboard schedules do not record who accessed what.

Restricting calls to in-platform endpoints and blocking external ones through policy enforces the boundary, regardless of what the application does. A system prompt instruction cannot control where data goes. Encrypting a prompt for an external provider still sends the data out, and storing prompts as comments does not limit exposure.