Databricks Generative AI Engineer Associate - Practice Exam 1

1. A claims pipeline must extract policy number, incident date, and claimed amount from varied document text into a predefined schema. Which task design is most suitable?
2. A company deploys a customer-facing agent UI. Which architecture keeps service credentials out of the browser?
3. A Databricks agent answers a follow-up question using only the words "what about last month?" as its retrieval query. The previous turn identified a specific region and product. What should the developer improve?
4. A Delta table stores parsed document chunks before indexing. Which fields best support incremental updates and traceable retrieval?
5. A developer changes the embedding model used for indexing a document corpus. Which two steps are necessary before trusting the new retrieval configuration? Select 2 answers.
6. A developer compares a retriever that finds relevant passages with a generator that ignores them. Which pair of metrics most directly distinguishes those behaviors?
7. A developer wants MLflow traces to distinguish retrieval delay from generation delay. How should the workflow be instrumented?
8. A document ingestion job stores extracted text from scanned inspection forms, but most rows are empty. What should the engineer check first?
9. A document source has a license that forbids the proposed commercial redistribution. What should the team do before adding it to a customer-facing RAG application?
10. A financial assistant must answer both policy questions from documents and numerical questions from governed tables. Which design best separates the two kinds of evidence?
11. A LangChain-style RAG chain has a retriever, prompt template, and model. What should be passed into the prompt template before generation?
12. A manufacturer wants an assistant that answers technicians' questions from maintenance manuals stored in Unity Catalog volumes. The team prefers a managed document-grounded solution. Which Agent Bricks capability should it evaluate first?
13. A model card states a candidate was trained primarily for English text and restricts certain commercial uses. What should be checked before using it for a multilingual commercial application?
14. A multi-tenant assistant searches a shared corpus. Which two controls should be part of its isolation design? Select 2 answers.
15. A nightly enrichment job must apply an LLM to a large Delta table, with results consumed the following morning. Which interface is appropriate to evaluate?
16. A policy assistant should answer only from approved evidence. A new prompt causes confident answers when retrieval returns nothing. What is the most targeted correction?
17. A procurement agent must find an approved supplier, check live stock, and create a draft purchase request. Which workflow ordering best respects the data dependencies?
18. A production agent must call a Databricks-managed MCP service and a separate external MCP service. What should be configured for each connection?
19. A RAG application receives user-provided document text that includes instructions to ignore the application policy. Which design is most appropriate?
20. A RAG corpus contains navigation menus repeated on every web page. Search repeatedly favors these menus over the article text. Which preparation is most useful?
21. A relevance-labeled retrieval test set has 8 relevant passages for one query. The top 10 results include 6 of them. What is recall@10 for that query?
22. A reranker is added after initial candidate retrieval. What role should it play?
23. A retrieval chunk contains relevant text but omits the product version. The answer generator combines it with instructions for another version. Which change is most useful?
24. A security review finds raw prompts and responses containing personal information in widely readable logs. Which remediation preserves necessary observability?
25. A service must send text to an external model but should not disclose customer account numbers. Which control is most direct?
26. A serving agent queries an AI Search index owned by another principal. Which Unity Catalog privileges are required for index access?
27. A standard AI Search Delta Sync index uses triggered synchronization. An ingestion job finishes successfully, but the index still serves the prior snapshot. What should the workflow add?
28. A team adds an MLflow scorer that compares responses with expected answers. Which evaluation dataset field is essential for that comparison?
29. A team needs a managed search index that incrementally reflects changes in an existing Delta table. Which AI Search index type fits?
30. A team selects a model for extracting short warranty clauses. Model A meets accuracy and latency targets at lower cost; model B is larger but offers no measured improvement. Which choice is best supported?
31. A team wants to prevent a prompt change from degrading known high-risk scenarios. Which release gate is appropriate?
32. An agent must answer current revenue questions using a Genie space. Why is a structured-data tool preferable to generating numbers directly from an LLM?
33. An agent must return a machine-readable object containing an integer severity and a list of evidence IDs. Which design most directly supports reliable integration?
34. An agent produces correct answers in development but slows down and costs more after launch. Which monitoring combination best supports investigation?
35. An agent retrieves the right source passages but produces statements inconsistent with them. Which intervention targets this stage of the failure?
36. An agent returns a wrong answer. A trace shows an incorrect tool argument even though retrieval and model serving both returned HTTP 200. What should the team investigate first?
37. An agent tool can cancel reservations. The model emits a syntactically valid cancellation call. What must happen before execution?
38. An application already computes embeddings outside Databricks and must directly upsert and delete vectors through its own ingestion service. Which AI Search index should it use?
39. An engineer must register a logged model in Unity Catalog. Which identifier format represents the target registered model?
40. An equipment corpus has precise section headings, warning blocks, and procedure steps. Random fixed-character splits separate warnings from the procedures they qualify. Which preparation change is most appropriate?
41. An MLflow pyfunc application works in a notebook but fails on a clean serving environment because a text-processing package is missing. What should be included in the logged model artifact?
42. An operations team must roll back a prompt after a regression while retaining version history. Which MLflow feature supports this workflow?
43. Before building a customer-service agent, which specification best translates the business goal into a testable AI application contract?
44. Experts reviewing the same agent answers disagree because "good" is undefined. What should the team do to make feedback useful for iterative evaluation?
45. Two prompts achieve similar correctness, but one returns twice as many tokens and often repeats source passages. What should the developer evaluate next?

Comments — 0