DB-GENAI Practice Test

DB-GENAI practice test: what's covered

• RAG
• Vector Search
• Mosaic AI
• LangChain
• Unity Catalog
• MLflow

Free DB-GENAI sample questions

1. A GenAI engineer is scoping a customer-support assistant that must answer from a 4,000-page internal policy PDF that changes weekly. The team wants accurate, up-to-date answers without retraining any model. Which application design best fits these requirements?

   • A. Retrieval-augmented generation that grounds a foundation model on retrieved policy chunks
   • B. Fine-tuning a base model on the policy PDF every week and serving the fine-tuned weights
   • C. Few-shot prompting that pastes the entire 4,000-page PDF into every request context
   • D. Pre-training a new transformer from scratch on the organization's policy corpus
   Show answer

   Correct: A. RAG retrieves the most relevant policy chunks at query time and grounds the LLM on them, so weekly content changes only require re-indexing, not retraining. Weekly fine-tuning is costly and slow to reflect changes, pasting 4,000 pages exceeds context limits and inflates cost, and pre-training from scratch is wholly disproportionate and would not stay current.

2. While decomposing requirements for a contract-analysis assistant, an engineer must map each business task to a model capability. The task 'pull the renewal date and counterparty name from a contract' is an example of which capability?

   • A. Information extraction of structured fields from unstructured text
   • B. Open-ended text summarization of the whole document
   • C. Sentiment classification of the contract's overall tone
   • D. Free-form creative generation of new contract clauses
   Show answer

   Correct: A. Pulling named entities such as a renewal date and counterparty is structured information extraction from unstructured text. Summarization compresses meaning rather than isolating fields, sentiment classification labels tone not facts, and creative generation invents new content, so none of those match the stated task.

3. An engineer is selecting evaluation and design constraints for a high-volume RAG chatbot that must keep latency low and cost predictable. Which TWO design choices directly help meet those non-functional requirements? (Choose TWO)

   • A. Cap the number of retrieved chunks and the generation max_tokens for each request
   • B. Select a smaller, lower-latency foundation model that still meets quality thresholds
   • C. Always retrieve the top 100 chunks to maximize recall for every query
   • D. Disable response caching so every identical question hits the model fresh
   • E. Send the full conversation history uncompressed on every single turn
   Show answer

   Correct: A and B. Bounding retrieved chunks and max_tokens limits prompt and output size, while a smaller model cuts inference time and cost; both directly serve latency and cost goals. Retrieving 100 chunks bloats the prompt, disabling caching wastes compute on repeats, and sending uncompressed history grows tokens every turn, so those choices work against the requirements.

4. A product team reports that their assistant frequently states confident but fabricated facts about company benefits. The engineer wants a primary mitigation that anchors responses to verifiable source material. Which approach most directly reduces these hallucinations?

   • A. Ground generation on retrieved authoritative documents and instruct the model to cite them
   • B. Raise the model's temperature so it explores more diverse candidate answers
   • C. Increase the maximum output length so the model has room to elaborate
   • D. Remove the system prompt so the model answers with fewer constraints
   Show answer

   Correct: A. Hallucinations drop when the model is grounded on retrieved authoritative context and asked to cite it, constraining answers to verifiable sources. Higher temperature increases randomness and makes fabrication more likely, a longer output gives more room to invent detail, and removing the system prompt removes guardrails, so each of those would worsen the problem.

5. An engineer is preparing long technical manuals for retrieval and notices that fixed 2,000-token chunks often split a single procedure across two chunks, hurting answer quality. Which chunking adjustment best preserves semantic coherence for retrieval?

   • A. Use smaller chunks with token overlap aligned to natural section boundaries
   • B. Increase chunk size to the model's full context window for every document
   • C. Concatenate the whole manual into one chunk so nothing is ever split
   • D. Randomly shuffle sentences before chunking to diversify the chunks
   Show answer

   Correct: A. Smaller chunks with overlap respecting headings and section boundaries keep each procedure intact and improve retrieval precision and recall. A full-context-window chunk dilutes relevance and raises cost, a single giant chunk defeats retrieval entirely, and shuffling sentences destroys the logical order needed to answer, so those are incorrect.

Last updated: 2026-10-19

DB-GENAI practice test FAQ