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 3 / Knowledge Retrieval and Genie Configuration
An internal knowledge agent answers employee questions from policy documents in a Databricks AI Search index, with hybrid search and a department filter already in place. Evaluation shows low recall@10 on questions whose wording differs from the documents: employees ask about 'PTO carry-over' when the policy says 'annual leave rollover', and about a 'laptop refresh' when the document says 'hardware replacement cycle'. The right documents are in the index and pass the department filter, but a log of every candidate the search retrieves, not only the top 10, shows that they never appear at any rank. Which change most directly improves recall on these questions?
Show answerHide answer
Answer: A
Why
The failure is vocabulary mismatch: the question and the passage say the same thing in different words, so the search never retrieves the right documents. The Databricks retrieval quality guide lists query expansion under Step 5, query optimization: generate multiple query variations, which 'improves recall by finding documents with different terminology'. Its example expands 'car maintenance' to 'automobile repair' and 'vehicle servicing'. The guide orders its steps from highest impact and lowest effort to lowest impact and highest effort, and hybrid search (Step 1) and filtering (Step 2) are already in place, but reranking (Step 3) evaluates documents that were retrieved, so it cannot fix a retrieval miss.
Why not the others
BThe reranker evaluates documents the search has already retrieved, and the guide presents it as a precision technique; it can reorder candidates but cannot add a document the search never retrieved.
CFull-text search matches on exact terms, so it fails when the question and the document use different words, and it gives up the semantic matching that hybrid search adds.
DA filter only shrinks the pool the search draws from; the right documents already pass the department filter and are still not retrieved, because the query does not match their wording.
Section 4 / Memory Architecture with Lakebase and MLflow
A mortgage pre-approval assistant on Databricks is written as a plain Python loop around a model serving endpoint; no agent framework manages its state. Within one conversation it collects an applicant's income, debts and property details over many turns, and after a restart of the app that conversation must resume with its ordered messages and tool results intact. Applicants return for several separate conversations over months, and each new one must recall durable facts they stated earlier, such as being self-employed and preferring a 15-year term, without asking again. To control cost, the context loaded when a conversation starts must stay small however many earlier conversations an applicant has had. Which store configuration fits?
Show answerHide answer
Answer: B
Why
The assistant needs both kinds of state Databricks describes. Agent sessions persist and resume an interaction's state, most commonly the ordered transcript of messages, tool calls and results, so that a conversation survives restarts. Agent memory stores durable facts, preferences and decisions that the agent recalls in later, separate conversations, retrieved by semantic search. Databricks says to use both together for most use cases, to use sessions alone when you only need in-interaction state, and to use memory alone when your framework already manages session state, which a plain loop does not.
Why not the others
ALoading every earlier transcript in full makes the starting context grow with each conversation, and the few durable facts that matter stay buried in it instead of being retrieved when they are relevant.
CMemory is scoped to a subject and retrieved by semantic search; it is not an ordered transcript of messages and tool results to replay, and Databricks lists memory alone for the case where your framework already manages session state, which this plain loop does not.
DA session is scoped to one interaction, so each new session starts without what earlier conversations produced; Databricks lists sessions alone for the case where you only need in-interaction state.
A facilities team is building an agent on Databricks that must create and update work orders in a third-party maintenance system through that system's HTTP API. The API credential must be managed centrally, agents from several teams will use the same tools, and access must be shared and restricted through Databricks governance. Which part of the Databricks stack provides this?
Show answerHide answer
Answer: C
Why
Databricks documents MCP as a way for an agent to call external services: you add a self-hosted or third-party server through the MCPs UI in Unity Gateway and share it with your team. Databricks-provided and registered MCPs use Unity Catalog permissions, and Unity Gateway is where you control access, manage credentials and monitor usage. That matches central credentials, tools shared across teams and governed access.
Why not the others
AAn agent memory store keeps durable facts for an agent to recall; it cannot create or update records in the maintenance system or manage the credential for its API.
BA description in the prompt gives the model no way to execute calls and no credential handling.
DA key hard-coded in the function body sits in source code instead of in a managed credential, such as a Databricks secret referenced by a Unity Catalog connection, so the central credential management the stem asks for is missing.
A procurement agent on Databricks calls a function that creates purchase requests. The function's parameter schema nests supplier, delivery address and line items four levels deep. MLflow traces show that the model often puts values at the wrong nesting level or omits required fields, and every call carries the full schema. Which change most improves the quality of the agent's tool calls?
Show answerHide answer
Answer: D
Why
Databricks documents that a simpler JSON schema for function call definitions results in higher quality function call JSON generation, that heavily nested JSON schemas result in lower quality generation, and that you should try flattening the schema for better results. Foundation Model APIs also accept only a subset of JSON schema: no anyOf, oneOf, allOf or $ref composition, and at most 16 keys in a schema. Flattening removes the cause of the misplaced values.
Why not the others
AWith 'required' the model always calls one or more functions; it does nothing about how the arguments are formed.
BAdds tokens and leaves the model facing the same deeply nested structure.
CThe arguments are misplaced, not cut off; more output room does not change the schema's shape.
Section 3 / Knowledge Retrieval and Genie Configuration
A company is building a policy-assistant agent on Databricks for all employees. Its chunk table holds parsed HR guidance, and some rows also carry, in an employee_id column, the ID of the employee whose case the guidance came from; a Unity Catalog column mask hides that column from everyone outside HR. Creating a Databricks AI Search index on the chunk table fails. Employees outside HR must never see the IDs, and the agent's answers do not need them. What should the team do?
Show answerHide answer
Answer: B
Why
The Unity Catalog page on row filters and column masks states: 'You cannot create an AI Search index from a table that has row filters or column masks applied.' The protection therefore has to happen before indexing, in the chunk table the index reads: writing it without the IDs keeps them out of the index and out of every context the agent builds, and the answers do not need them. AI Search also documents that 'Row and column level permissions are not supported', so the index itself cannot hide a column from some users.
Why not the others
AThe limitation applies to any table that has a column mask applied, so an exception inside the mask function does not make the table indexable; reading the raw IDs would also copy them into the index.
CA row filter blocks index creation in the same way as a column mask, and it would also withhold guidance that employees outside HR need.
DWithout the mask the IDs go into the index and can reach any employee's context; an instruction in the prompt is advisory and does not keep them out.
The study bundle has two practice exams with 110 original questions, an answer key and an explanation for every answer, plus the 63-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.