Databricks Certified Generative AI Engineer Associate (DB-GENAI) Practice Questions and Study Guide

A downstream service parses the output of your LLM step and fails whenever the model adds a friendly sentence before the data. The service expects an object with the keys "category" and "priority". What is the most reliable prompt design?

• AAsk the model to be concise and avoid any unnecessary words or filler in its answer
• BGive the exact JSON schema, one example output, and say to return only JSON
• CRaise the temperature so the model explores more output formats
• DAsk the model to put the data somewhere near the end of its answer

Your source is a set of technical manuals with clear headings and sections. Answers are often lost because fixed size chunks cut sections in half. Which chunking approach should you try first?

• AUse fixed size chunks of exactly one sentence each with no overlap
• BEmbed each whole manual as one single chunk without any splitting
• CSplit the text every fixed number of characters, ignoring any headings
• DSplit on document structure such as headings, then cap the chunk size

A user asks "What is the return window for my order 4471?" Before retrieval, you want the prompt to include facts about that order so the answer is specific. What is this technique?

• AAugmenting the prompt with context looked up from key fields in the input
• BFine tuning the model on every order record the company has ever had
• CLowering the max tokens setting so the model answers in a single short sentence
• DRemoving the order number from the question to protect user privacy

You need to run sentiment classification with an LLM over 20 million rows in a Delta table every night. What is the most appropriate approach on Databricks?

• ACall the real time chat endpoint row by row from a Python loop in a notebook
• BSync the table into a Vector Search index and query it once per row nightly
• CUse ai_query() in a SQL or Spark job to run batch inference on the table
• DExport the table to CSV and paste the rows into a chat window

Support tickets contain customer email addresses and phone numbers. You want a RAG assistant to use the ticket text but never reveal those details. What should you do first?

• ATell users in the chat window not to ask for any personal details
• BMask or redact personal details in the source before chunking and indexing
• CUse a larger LLM that is better at ignoring sensitive information
• DIndex the raw tickets as they are and hope the retriever skips sensitive chunks