Independent: not affiliated with Databricks. Free to try; the study bundle adds a study guide and two practice exams.
5 original questions, none of them in the practice exams.
Pick an answer. Show answer explains every option and links its Databricks source.
5 questions / Explanations link Databricks sources / Last updated Oct 26, 2026
Question 1
Multiple choice
Section 1 / Design Applications
A marketplace lands about 50,000 product reviews a day in a Delta table. Reviews arrive in Spanish, German, French and Italian; the analysts who read them work in English. Within an hour of arrival, each new review's full text must be stored in English in a second column, with the original kept. Which model task, and which Databricks AI Function, meets the requirement?
Show answerHide answer
Answer: D
Why
The requirement is the same text in another language, which is a translation task. ai_translate(content, to_lang) invokes a model to translate text to a specified target language from SQL; to_lang accepts BCP 47 codes such as 'en', and English, French, German, Italian and Spanish are among the languages the function page lists. Running it over the new rows on a schedule writes the English column without a person in the loop.
Why not the others
AA summary condenses the review; the requirement is the full text in English, which summarization does not produce.
BSentiment analysis returns a label such as positive or negative, not the review's text.
Cai_similarity compares two strings and returns a similarity score; it does not produce a translation.
A product team is building a RAG assistant over 3,000 user-research interview transcripts stored as plain text in a Unity Catalog volume. The transcripts have no headings and no paragraph breaks, and a single interview often moves from onboarding to pricing to a competitor's product. With fixed 512-token chunks, many retrieved chunks join the end of one topic to the start of the next, and answers pick up details from the wrong topic. Which chunking strategy fits these documents?
Show answerHide answer
Answer: B
Why
The Databricks guide to building an unstructured data pipeline for RAG describes semantic chunking: 'Techniques such as topic modeling can be applied to identify semantically coherent sections in the text.' These transcripts have no structure to split on, and the failure is chunks that span a change of topic, so boundaries placed where the topic changes give each chunk a single topic.
Why not the others
AThe guide warns that larger chunks 'can include irrelevant information'; with boundaries still at fixed token counts, a 2,048-token chunk spans more changes of topic, not fewer.
CThe guide's format-specific chunking uses the structure that formats such as Markdown or HTML carry; these transcripts have no headings to split on.
DOverlap repeats text across chunk borders, but the borders still fall at fixed token counts, so chunks keep mixing topics; 50% overlap also creates near-duplicate chunks.
A travel-booking agent implemented as a ResponsesAgent receives, with every request, a client_type and a session_id that are not part of the conversation. The agent also retrieves source links for each answer; the client must display them, but they must stay out of the chat history used on later turns. Which two steps meet these needs? Choose two.
Show answerHide answer
Answers: C and D
Why
MLflow ResponsesAgent natively supports the fields custom_inputs and custom_outputs for scenarios that need additional agent inputs, such as client_type and session_id, or outputs like retrieval source links that should not be included in the chat history for future interactions. The agent reads the extra inputs from request.custom_inputs and returns the links in custom_outputs, so neither one travels in the conversation messages.
Why not the others
AText appended to the assistant message becomes part of the chat history, which the requirement rules out; the page puts outputs such as retrieval source links that should stay out of the history in custom_outputs.
BTags on a registered model version are set once for the version, so they cannot hold values that change from one request to the next.
EMessages in request.input are the conversation, so a prefix there becomes part of it; the page puts additional inputs such as client_type and session_id in custom_inputs.
A Python service calls an outside vendor's hosted chat model through the OpenAI Python client's chat completions interface. The company wants the service to use a Databricks-hosted chat model on a Foundation Model APIs pay-per-token endpoint instead, with the smallest change to the service's code. Which approach does Databricks document for this?
Show answerHide answer
Answer: C
Why
Model Serving exposes a unified OpenAI-compatible API and SDK for querying foundation models. The documentation lists the OpenAI client as a query method for chat, embeddings and completions models made available by Foundation Model APIs or external models, with the serving endpoint name given as the model input; the databricks-openai package provides an OpenAI client with authorization automatically configured. The service keeps its client code and changes the model name and the credentials.
Why not the others
Aai_query() is the SQL function for inference from queries; rewriting a Python service's calls as SQL statements is a larger change than keeping the client.
BFoundation Model APIs pay-per-token endpoints already serve the Databricks-hosted model; wrapping it in a custom pyfunc endpoint adds a deployment the requirement does not need.
DAn external model endpoint proxies the third-party provider's model; the requirement is a Databricks-hosted model on Foundation Model APIs.
A team evaluates a support agent with a custom field-based judge built with make_judge(), named resolution_quality, that returns pass or fail from each trace's inputs and outputs. Four support leads reviewed 120 traces and logged their own pass or fail verdicts with mlflow.log_feedback under the name resolution_quality; about a third of their verdicts are fail. The judge disagrees with the leads on roughly a quarter of those traces. Before the judge is used in production monitoring, the team wants it to apply the leads' standard. What should they do?
Show answerHide answer
Answer: D
Why
Judge alignment teaches an LLM judge the human standard, and the same align() workflow applies to custom judges from make_judge(). The leads' feedback is logged under a name matching the judge's, on far more than the 10-trace minimum, with over 30 percent negative examples; align() returns a new, aligned judge.
Why not the others
ACorrectness compares a response with ground truth expectations, which the leads did not write; their pass or fail verdicts are feedback, not expectations, so it would not apply their standard.
BThe new judge reads the same instructions and never sees the leads' verdicts, so nothing in this change carries their standard into it; align() is the step that takes the verdicts as input.
CThe page's best practices call for at least 30 percent negative examples, and this choice would leave none; without the fail verdicts the judge never sees what the leads rejected.
The study bundle has two practice exams with 110 original questions, an answer key and an explanation for every answer, plus the 62-page study guide. Practice exam explanations are shorter than the ones on this page: a short paragraph per question, with no source links.
Study bundle
$19
110 practice questions / Study guide / Answer keys
Every bundle includes 6 months of access: download the PDFs, and any new editions, until 6 months after your purchase date. Your receipt has a license key. Enter it at overqual.com/download for the three PDFs, stamped with your email and order number. They are yours to keep.