Databricks Certified Generative AI Engineer Associate

Correct answer: B. Sentiment analysis -> NLU, Summarization -> NLG, Translation -> NLG, Dialogue -> NLG

Sentiment analysis is natural language understanding because it interprets input text, while summarization, translation, and dialogue all produce new text, so they fall under natural language generation. Remember the split: NLU covers classification and extraction tasks; anything that writes output is NLG.

Correct answer: B. Refine and optimize the prompt with detailed instructions using prompt-engineering techniques so the model has enough context for the task.

When the model and tokenizer work but outputs are vague, the fix is better prompt engineering: add context, constraints, format, and few-shot examples. Increasing parameters or switching architectures is irrelevant when the prompt itself lacks the needed guidance.

Correct answer: B. Building a support chatbot that must fetch the latest product policies and give accurate, real-time answers.

RAG shines when the chatbot must fetch current, private documents like product policies and answer with accurate grounding. Pure generation from internal training data needs no retrieval, so those scenarios miss the point of RAG.

Correct answer: A. Embeddings turn text into vectors, the vector database stores those vectors, search and retrieval find the most relevant ones, and prompt augmentation merges them with the user query.

The canonical RAG flow is embed, store in a vector database, retrieve the most similar vectors, then augment the prompt with that content before the LLM call. Any option that swaps these roles, like vector databases generating embeddings, is wrong.

Correct answer: C. Pass relevant context documents into the LLM prompt during query processing.

Passing relevant policy documents into the prompt, the core of RAG, grounds the model's answers and reduces hallucination. Jumping straight to fine-tuning is the classic distractor; grounding context is the first-line fix, not weight updates.

Correct answer: A. It stores embeddings and enables fast semantic search by finding vectors with similar meaning rather than relying only on exact keyword matches.

A vector store indexes embeddings and enables fast semantic similarity search, so related meanings match even without shared keywords. Note that it does not create embeddings itself; embedding models do that before storage.

Correct answer: B. It balances context retention and retrieval precision by splitting documents into appropriately sized, semantically coherent pieces.

Good chunking splits documents into right-sized, semantically coherent pieces, balancing enough context against precise retrieval. Fixed-size chunks with no overlap often cut ideas in half, and chunking never replaces the embedding step.

Correct answer: B. The system retrieves relevant policy documents from the vector database, combines them with the user query into an augmented prompt, and passes it to the LLM, which produces a grounded answer.

RAG retrieves the relevant internal documents from the vector store, augments the user query with them, and the LLM generates a grounded answer. Fine-tuning the documents into weights is the key distractor; RAG fetches at query time instead.

Correct answer: A. Different embedding spaces – queries and documents were embedded with different models, producing mismatched vector spaces.

Embeddings from different models live in incompatible vector spaces, so similarity comparisons become meaningless and results are irrelevant. The other options are tuning issues that would not cause consistent irrelevance like a mismatched embedding model does.

Correct answer: A. LangChain

LangChain is purpose-built for LLM workflows with prompt templates, chains, agents, memory, and RAG retrieval integration. Spark, TensorFlow, and Delta Lake serve data processing, deep learning, and storage rather than orchestration.

Correct answer: C. It optimizes latency, since reranking makes the query return faster.

Reranking adds an extra scoring step, so it increases latency rather than reducing it; its real value is retrieval quality. Watch for NOT questions where a speed claim is the false benefit.

Correct answer: A. To merge the user's original query with retrieved context, giving the LLM in-context learning for accurate, source-backed answers.

Prompt augmentation combines the user's query with retrieved context so the LLM can answer via in-context learning. The retraining option fails because RAG avoids retraining and instead grounds answers in retrieved sources.

Correct answer: A. It uses zero-shot classification, assigning labels from its general knowledge.

With no labeled examples in the prompt, the model applies zero-shot classification, mapping inputs onto the given labels from its pretrained knowledge. Few-shot would require example labelings in the prompt, so that distractor does not fit here.

Correct answer: D. Vector search finds semantically relevant documents by comparing embedding similarity, returning results based on meaning rather than exact keyword matches.

Vector search compares embedding similarity so retrieval matches meaning rather than exact keywords, which is exactly what option D states. The distractors describe generation, raw-text matching, or date ranking, none of which reflect how similarity search works.

Correct answer: A. Convert the user query to an embedding with the same model used for the chunks, compute similarity against stored document embeddings, rank and retrieve the top-k chunks, then assemble them with the original question for the LLM. It beats keyword search because embeddings capture meaning (e.g., 'canine care' is close to 'dog training').

The retrieval flow is query embedding with the same model, similarity computation, top-k ranking, then context assembly for the LLM. Using one embedding model keeps query and document vectors in the same space; the other options skip retrieval entirely.

Correct answer: C. A vector database stores embeddings and metadata and uses specialized indexes to retrieve vectors by mathematical similarity rather than exact matches, integrating with RAG for semantic search.

A vector database stores embeddings plus metadata and uses specialized indexes to retrieve by mathematical similarity, enabling semantic search in RAG. Options claiming it removes embeddings or replaces retrieval misstate its core purpose.

Correct answer: A. To re-evaluate retrieved chunks beyond raw similarity scores and order them by true relevance to the query before sending them to the LLM.

Reranking re-scores retrieved chunks for true relevance to the query, improving the ordering beyond raw similarity scores. Options about generating embeddings or fine-tuning the LLM describe other pipeline stages, not reranking.

Correct answer: A. When your dataset is small enough to fit in memory, changes rarely, and you need fast similarity search without deploying a full database.

FAISS suits small, static, in-memory datasets where you need fast similarity search without a deployed database. Large, dynamic datasets needing real-time updates and multi-user access call for a full vector database instead.

Correct answer: B. Evaluate chunking, the embedding model, the vector store, retrieval/re-ranking, and the generator – both individually and as part of the full pipeline.

Proper RAG evaluation covers every component, including chunking, embeddings, the vector store, retrieval, reranking, and the generator, both individually and end-to-end. Evaluating only the generator ignores upstream retrieval failures that degrade final answers.

Correct answer: B. Use an OAuth 2.0 flow with token issuance and automatic rotation.

OAuth 2.0 with short-lived, automatically rotated tokens minimizes credential exposure and supports revocation. On the exam, remember that hardcoded or shared static keys are always the wrong answer for production security.

Correct answer: B. Create a separate Genie Space for each task agent, scope each space to only that domain's Unity Catalog tables, and run each agent's serving endpoint under its own service principal that is granted UC access (USE CATALOG, USE SCHEMA, SELECT) only to its own data; the supervisor agent routes each request to the correct subagent.

Genie Space data access is enforced by Unity Catalog permissions tied to the running identity, not by the space itself. Prompt instructions and shared service principals are classic distractors because they provide no real isolation.

Correct answer: A. Sentence-based chunks

Sentence-based chunks create focused, high-precision embeddings ideal for short, specific questions. Paragraph chunks dilute meaning for narrow queries, so match chunk size to the granularity of the questions.

Correct answer: B. It means the application forwards the end-user's identity to downstream systems so access policies apply consistently, achieved via SSO tokens or delegated credentials.

Authentication passthrough forwards the end user's identity via delegated SSO or OAuth tokens so downstream permissions still apply. A shared service identity defeats the purpose because all queries bypass per-user access rules.

Correct answer: A. Cosine similarity

Cosine similarity is the standard measure for semantic closeness between embeddings because it is scale-invariant and direction-focused. ROUGE, BLEU, and Levenshtein compare surface text, so they cannot capture embedding-level semantics. On the exam, match similarity metrics to vector retrieval, not text generation.

Correct answer: B. Apply a metadata filter at retrieval so only documents with event_date BETWEEN '2002-03-27' AND '2002-04-05' are included.

Applying a metadata filter at retrieval guarantees only documents inside the five-day window are returned. Relying on the LLM or cosine similarity to filter dates is unreliable, so filter structured attributes before generation on the exam.

Correct answer: B. Unstructured

The Unstructured library parses HTML, PDFs, and images with OCR support, making it ideal for extracting clean text before chunking. BeautifulSoup only handles HTML, and NLTK and TensorFlow do not perform document extraction.

Correct answer: B. Return only valid JSON with keys: customer_id, order_id, order_date, order_status. Do not include any text or explanation outside the JSON.

The best prompt explicitly lists the required JSON keys and forbids any text outside the JSON, making output deterministic and parseable. Vague instructions like 'answer in JSON format' leave room for extra prose and unpredictable structure.

Correct answer: C. For each row, use regex to strip specific details (file name, IDs, etc.) to sanitize the message, then apply ai_summarize() to summarize the sanitized text.

Sanitizing the messages with regex first removes unique details, so ai_summarize() can group variants under one root-cause category. A generic summarize prompt without normalization keeps each message distinct and prevents clean categorization.

Correct answer: A. Use Auto Loader with a custom UDF to split the PDF text into chunks, then explode the array to flatten it into individual rows.

Auto Loader with a UDF provides scalable ingestion and chunking, and exploding the array produces one chunk per row for embedding. Looping outside Spark or storing whole documents breaks scalability and vector-search compatibility.

Correct answer: B. Ingest documents -> Index to Vector Search -> User submits query -> LLM retrieves documents -> LLM generates a response -> Evaluate the model -> Deploy with Model Serving

The correct flow ingests documents, indexes them to Vector Search, retrieves relevant content for the query, generates the answer, then evaluates and deploys. Evaluation must happen after generation, which is why options evaluating before generation fail.

Correct answer: B. Add explicit instructions that the model must answer only from the provided input and must say 'I don't know' when information is missing.

Grounding the model to answer only from provided input and admit 'I don't know' directly reduces hallucination and fabrication of private data. Encouraging detailed output despite missing data is the main trap and increases hallucination risk.

Correct answer: A. Document embeddings build a searchable knowledge base during preparation, while query embeddings convert each user query into the same vector format for real-time semantic comparison.

Document embeddings build the searchable knowledge base at preparation time, while query embeddings convert user questions into the same vector space at runtime. Remember both must come from the same embedding model for comparison to work.

Correct answer: C. Retrieve shipping data from an authoritative transactional data source (for example, Lakebase, a Delta table, or an operational database) keyed by transaction_id, and pass it to the LLM at query time.

Shipping status and arrival dates are live operational data, so retrieving them from an authoritative transactional source keyed by transaction_id keeps answers accurate. Assuming a fixed delivery window or indexing static policies cannot reflect changing shipping facts.

Correct answer: C. pytesseract

pytesseract wraps the Tesseract OCR engine, so it extracts text from scanned images in just a couple of lines. The other options are HTML parsing or scraping libraries and cannot read text embedded in image files.

Correct answer: D. Summarization

Summarization condenses a paragraph into a shorter sentence while preserving intent, which matches the goal. Text2text generation is the broader category, so picking the specific task gives the best model evaluation fit.

Correct answer: B. Matching the model to your data properties, domain, and content type.

The embedding model should match your data's domain, content type, and vocabulary because alignment drives retrieval quality. Picking the largest or cheapest model ignores semantic fit, and user count does not determine embedding choice.

Correct answer: D. All of the other options are correct.

Long contexts raise cost per request, increase inference latency, and trigger the lost-in-the-middle effect, so all listed challenges apply. Remember that when every individual option is valid, the comprehensive all-of-the-above choice is correct.

Correct answer: A. Use an MLflow pyfunc to package the LLM workflow into a model while implementing retry and exponential backoff in the application or model code.

Packaging the workflow as an MLflow pyfunc model with retry and exponential backoff handles throttling robustly while keeping versioning and deployment. Temperature, random scheduling, and disabling tracking do not address rate limits at all.

Correct answer: C. Fine-tune a custom embedding model, use standard vector search, and keep hybrid search and reranking off.

A custom embedding model improves accuracy, standard vector search has the lowest latency, and disabling hybrid search and reranking avoids added compute. Storage-optimized endpoints trade latency for cost, so they fail the latency requirement.

Correct answer: C. The request is blocked and the client receives a default blocked response. The event is recorded in usage tracking, and guardrail activity can be audited through the evaluator endpoint if inference logging is enabled.

A blocking PII guardrail rejects the request before it reaches the model, returning a default blocked response to the client. Blocked events still appear in usage tracking, but the endpoint inference table only holds requests that reached the model.

Correct answer: A. Context length 512: smallest model 0.13 GB, embedding dimension 384.

A context length of 512 exactly matches the chunk size, so the smallest model with the lowest latency and cost suffices. Larger context windows would waste resources since chunks never exceed 512 tokens.

Correct answer: F. It learns from the few in-prompt examples and predicts the sentiment from context.

Three labeled examples inside the prompt make this few-shot, in-context learning, so the model predicts Positive for the fourth tweet with no retraining. Zero-shot is the trap; it only applies when no examples are supplied in the prompt.

Correct answer: G. formatted_prompt = template.format(document="This is my text")

PromptTemplate.format requires a value for every declared input variable, so you must pass document as a keyword argument. Omitting it or adding undeclared variables like model raises a missing-variable error.

Correct answer: B. Combine human evaluation with automated metrics to assess tone, appropriateness, and contextual relevance.

The correct answer works because combining human review with automated metrics captures subjective qualities like tone and empathy that similarity scores miss. On the exam, remember that automated metrics alone cannot judge helpfulness or appropriateness of customer-facing responses.

Correct answer: B. Evaluate the whole AI system – the LLM plus its integration with retrieval, UI, and business workflows – and the overall user experience.

The correct answer works because end-to-end evaluation covers the full AI system, including retrieval integration, user experience, and business impact. The distractor about LLM-only benchmarks fails because it ignores the integrated components that determine real-world success.

Correct answer: B. Create MLflow evaluation records with example queries, define a metric object with relevance criteria and scoring scale, and use an LLM judge to score each response.

The correct answer works because the MLflow workflow is to build evaluation records, define a metric with criteria and a scoring scale, then have the LLM judge score responses. Loss and perplexity do not measure relevance directly, which is why that distractor fails.

Correct answer: B. Create a feature store table keyed by order_id, populated with tracking data, courier details, and estimated delivery dates.

The correct answer works because a feature store keyed by order_id supplies accurate, live, order-specific tracking and delivery data to the app. The vector store distractor fails because shipping status needs real-time data, not static policy documents.

Correct answer: B. Contact the data curator or data governance team to validate whether the dataset is approved and compliant.

The correct answer works because escalating to the data curator or governance team validates dataset approval and compliance before further use. Exam cue: never silently delete, keep using, or trust an LLM to filter PII; always route governance issues to the responsible team.

Correct answer: C. Feature Serving

Feature Serving is the correct choice because it exposes fresh online features to applications with low latency via a serving endpoint. Foundation Model APIs serve models, not feature data, so it does not deliver the live scores needed here.

Correct answer: B. vsc.create_delta_sync_index()

A Delta sync index automatically stays synchronized with the source Delta table as rows are inserted, updated, or deleted. A direct vector access index must be synced manually, so it does not meet the auto-update requirement.

Correct answer: D. Add few-shot examples showing the desired summary style without explanations.

Meta-commentary is a prompt formatting problem, so few-shot examples of clean summaries teach the desired output style most reliably. Changing chunking or embeddings targets retrieval quality, which the question says is already fine.

Correct answer: D. Assess and improve the quality of the retrieved context.

RAG accuracy failures usually come from retrieval, so evaluating and improving the retrieved context is the priority. Swapping the LLM treats generation when irrelevant documents are the actual root cause.

Correct answer: D. Wrap the workflow in a custom MLflow PyFunc model (a class inheriting mlflow.pyfunc.PythonModel) and perform the prompt preprocessing inside predict() before it calls the LLM, with post-processing applied to the output afterward.

A custom MLflow PyFunc model packages preprocessing, the LLM call, and post-processing into one deployable asset, with prompt preprocessing happening inside predict() before the LLM is called. Hardcoding a prompt cannot adapt to per-request input, and post-processing alone runs too late to shape the prompt.

Correct answer: D. When the goal involves several intents or tasks (e.g., translation, summarization, sentiment), breaking it into ordered stages improves reasoning and reliability.

The correct answer works because decomposing a multi-intent goal into ordered stages keeps each prompt focused and feeds outputs forward, improving reliability. A practical cue: watch for distractors claiming multi-stage systems remove external tools or run parallel calls for speed; those are fabricated trade-offs.

Correct answer: C. LlamaIndex

LlamaIndex is correct because it is the data-centric framework purpose-built to index SQL databases, APIs, and documents for retrieval. Remember the pairing: LlamaIndex for data ingestion and retrieval, LangChain for general chaining and orchestration of multi-step workflows.

Correct answer: B. LangChain – general-purpose LLM apps with chains, retrievers, tools, and flexible integrations.

LangChain fits because it supports chaining multiple LLM tasks plus retrievers, vector stores, and external tools with high flexibility. The strongest distractor, LlamaIndex, is great for data connections but is not as comprehensive for orchestrating complex multi-task workflows.

Correct answer: C. Chain to orchestrate multi-step tasks, Prompt to guide the LLM, Retriever to find relevant context, and Tool to integrate external functionality such as APIs.

The correct option maps each component properly: Chain orchestrates, Prompt guides the model, Retriever fetches context, and Tool integrates external APIs. The distractors simply scramble these roles, so match each component to its function rather than memorizing order.

Correct answer: B. Query Engine – indexes, stores, and queries documents to retrieve relevant chunks for RAG.

The Query Engine is correct because it indexes the knowledge base and retrieves the most relevant context for user questions. The Agent distractor handles multi-step autonomy but is not the primary retrieval interface for straightforward RAG questioning.

Correct answer: B. Agent – an autonomous entity that can read, write, run reasoning loops, adapt, and collaborate with other agents.

The Agent is right because it can read and write data, run reasoning loops with feedback, and collaborate with other agents. Cue: whenever a scenario mentions real-time writes or autonomous multi-step behavior, choose Agent over Query Engine.

Correct answer: A. Retriever to fetch relevant documents, Generator to create responses with LLMs, and Pipelines to orchestrate the flow.

The correct combination is Retriever to fetch relevant documents, Generator to produce answers, and Pipelines to orchestrate the flow. The distractors swap component roles, so anchor each Haystack component to its single core job.

Correct answer: A. Use Signatures to specify desired behavior, Modules to implement them with prompting techniques, and Optimizers to automatically improve performance.

Signatures declare the desired input-to-output behavior, Modules implement them with prompting techniques, and Optimizers automatically refine the prompts. Cue: remember the DSPy trio as declare, implement, optimize; distractors misassign roles like embedding generation to Signatures.

Correct answer: A. LLM-as-a-judge uses another LLM to evaluate model outputs with clear prompts, scoring scales, and structured formats for consistent, scalable evaluation.

The correct answer works because LLM-as-a-judge uses another LLM with structured prompts and scoring scales to evaluate outputs consistently at scale. Remember, it never replaces evaluation with fine-tuning or auto-selection; it requires explicit evaluation prompts and rubrics.

Correct answer: D. Summarization

The correct answer works because summarization condenses longer text into a shorter version that preserves the core meaning and intent. Generic text-to-text generation is the strongest distractor, but it is too broad for the specific condensing use case described.

Correct answer: D. Inference Tables

Inference Tables automatically log all endpoint requests and responses into a Unity Catalog Delta table, removing the need for a custom logging layer. Vector Search handles similarity lookups and does not capture endpoint traffic.

Correct answer: C. vsc.get_index(index_name)

get_index retrieves the metadata, configuration, and status of an existing Vector Search index. The create functions build new resources, and similarity_search runs queries rather than inspecting index details.

Correct answer: C. Use llm_chain = LLMChain(llm=OpenAI(), prompt=prompt) and call llm_chain.run({"adjective": "funny"}).

The chain needs a real LLM instance passed as llm, and run executes a single input through the chain. The original code omits the model entirely, so it fails no matter how generate is called.

Correct answer: A. Use a neutralizer to normalize the tone and style of the underlying documents.

Neutralizing the source documents strips the expressive tone the creative task depends on, so it hurts stylistic alignment. Few-shot examples, explicit tone instructions, and fine-tuning all directly steer output style.

Correct answer: A. Log the model with MLflow during training, register it in Unity Catalog via the MLflow API, and start a Model Serving endpoint.

Logging with MLflow, registering in Unity Catalog, and enabling Model Serving is the native Databricks deployment path. The Docker and Flask options work in general but abandon platform integration, so they fail the 'simplest and most integrated' requirement.

Correct answer: A. Create an agent that uses specialized tools – one to query Delta tables for prices and another for live web search – and feed both contexts to the LLM.

An agent with a SQL tool for Delta table prices and a web-search tool for live news supplies current context to the LLM. The pure RAG option fails because a vector store can't guarantee real-time news or fresh prices at query time.

Correct answer: B. Add credentials as environment variables, securely injected into the serving environment.

Environment variables sourced from Databricks Secrets are the recommended way to inject credentials into a Model Serving PyFunc environment. Spark session configuration doesn't apply to serving, and plaintext secrets in code or requests are insecure.

Correct answer: D. Select a smaller, domain-specialized LLM fine-tuned for biomedical/scientific content.

A smaller domain-specialized model tuned for biomedical content gives strong relevance at lower inference cost. The largest-model option maximizes cost, while limiting retrieval or queries degrades quality instead of improving efficiency.

Correct answer: B. Track prompts as MLflow versions and promote them using aliases after they pass.

MLflow prompt versioning with aliases provides gated promotion, full history, and one-step rollback to any prior version. Git branching also versions prompts, but it lacks the platform-native alias-based promotion and evaluation integration the exam expects.

Correct answer: A. Use a Databricks App backend to call the Agent endpoint with the app's credentials and enforce user identity/permissions via the app's authenticated context.

A backend-for-frontend pattern keeps credentials server-side and passes the authenticated user context so per-user permissions are enforced. Embedding a PAT in browser JavaScript or an API key in the frontend exposes tokens to any client.

Correct answer: B. Define clear rubrics, calibrate the SMEs on the criteria, and use the aligned judgments in mlflow.genai.evaluate() for consistent evaluation.

Defining clear rubrics and calibrating the subject-matter experts aligns their judgments, producing consistent labels usable in mlflow.genai.evaluate. Simply averaging disagreeing scores or discarding disputed cases hides inconsistency rather than fixing it.

Correct answer: B. Use Databricks Lakebase – a fully managed, Unity Catalog-governed Postgres (OLTP) database – to store the chat, session, and message IDs.

Databricks Lakebase is a managed Postgres OLTP database that provides ACID transactions, low-latency row-level access, and Unity Catalog governance. Delta tables excel at analytics but are not optimized for fast single-row reads and writes per interaction.

Correct answer: C. Run batch inference with the ai_query() SQL function against a pay-per-token Foundation Model API directly over the Delta table.

The ai_query() function applies a foundation model to a Delta table column in one SQL statement with pay-per-token pricing and no always-on endpoint cost, which is ideal for a small batch. A real-time endpoint adds idle cost that a few hundred rows cannot justify.

Correct answer: B. LangChain – chains multiple LLM calls with built-in support for retrievers, tools, prompts, and integrations to simplify complex GenAI apps.

LangChain is the intended answer because it abstracts the wiring between chains, retrievers, prompts, and tools. However, the scenario heavily overlaps with the previous question, so both LangChain and Haystack pipelines could be defended depending on emphasis on integration breadth.

Correct answer: B. Configure embeddings for semantic search, use a Retriever to find relevant chunks, build a Chain that combines retrieval with answer generation, and optionally add a second Chain to create the social post.

The staged design works because embeddings and a retriever ground the answer in the policy documents, and a chain links retrieval to generation, with an optional second chain for the post. The single-prompt option fails because it skips retrieval and risks hallucination.

Correct answer: B. Measure perplexity to understand the model's uncertainty in predicting the next token.

The correct answer works because perplexity is a standard objective metric that reflects the model's uncertainty when predicting the next token. The other options describe training or infrastructure actions, not evaluation objectives, so they fail the question's scope.

Correct answer: A. Context length 256, smallest model 0.12 GB, embedding dimension 256.

The correct answer works because matching context length to the 256-token chunks minimizes cost and latency, which the scenario prioritizes. Larger options like 1024 or 8192 waste capacity and increase cost when long conversations are not a requirement.

Correct answer: A. CodeLlama-34B

CodeLlama-34B is purpose-built for code generation and completion across many programming languages. Llama-2-70B is a general-purpose model, and BGE-Large produces embeddings rather than generating code.

Correct answer: C. Chunk into 512-token chunks with a 50-token overlap.

Chunks at the maximum 512 tokens with a 50-token overlap preserve continuity across boundaries in long legal contracts. The 256-token option wastes context capacity, though a test-taker could argue smaller chunks improve retrieval precision, making the wording slightly debatable.

Correct answer: A. Diversity of responses

With accuracy and relevance already handled, response diversity is what keeps long conversations fresh and engaging. Randomness is uncontrolled variety and can hurt coherence, making it the strongest distractor to avoid.

Correct answer: C. Deploy Llama-2-70B on Databricks Model Serving using Foundation Model APIs (with provisioned throughput), governed by Unity Catalog.

A large open model served inside Databricks under Unity Catalog governance keeps sensitive data in-house and maximizes accuracy. The external GPT-4 option sends data outside the governed environment, which directly violates the requirement.

Correct answer: B. Zero-shot gives no examples, while few-shot provides a small number of examples to guide behavior.

Zero-shot prompting provides no examples, while few-shot prompting includes a handful of demonstrations to steer the model. Any option claiming zero-shot supplies examples reverses the definitions and is wrong.

Correct answer: C. When the task involves multi-step reasoning and you want the model to show intermediate steps before the final answer.

Chain-of-thought prompting is used for multi-step reasoning where the model shows intermediate steps before answering. Option D describes prompt chaining, which splits tasks into sequential sub-prompts rather than reasoning step by step.

Correct answer: C. Use hierarchical chunking that preserves multi-level context (section -> subsection -> paragraph), embedding both parent and child text.

Hierarchical chunking preserves section structure and embeds both parent and child text, so retrieval captures broad and fine-grained context for long multi-section documents. The distractor about removing overlap cuts cost but reduces recall, which typically worsens relevance.

Correct answer: C. Review the dataset's license and terms of use before using it, and confirm your intended use (training, RAG, commercial deployment) is permitted, securing the necessary rights up front.

Reviewing the license and terms before use, and confirming training and commercial deployment are permitted, is the correct way to manage legal risk. Waiting until after deployment or only reacting to complaints exposes the organization to violations and does not secure the needed rights.

Correct answer: C. Deploy the RAG agent as a Databricks App and configure SSO with the organization's Microsoft Entra ID (Azure AD) so analysts sign in with their existing Office 365 identity, then surface the app through the Microsoft Teams they already use.

A Databricks App with SSO federated to the existing Microsoft Entra ID lets analysts use their current corporate identity with no new licensed application. Notebook-based access defeats the purpose of an assistant interface and adds workspace administration overhead.

Correct answer: B. Model B, for its higher safety and lower latency.

Model B is correct because the organization prioritizes safety and policy compliance, where B leads while still keeping strong accuracy and lower latency. Option D is also defensible, but B names safety and latency, the stated priorities.

Correct answer: D. Check availability, embed all employee profiles into a vector store, and use the project scope as the query with availability filters to retrieve the best matches.

Embedding employee profiles once into a vector store and querying with the project scope, plus availability filters, gives scalable semantic matching. Brute-force similarity over every profile doesn't scale for a very large team.

Correct answer: B. Deploy with on-demand, pay-per-token pricing that optimizes cost for low volume.

On-demand pay-per-token pricing is correct because it charges only for actual usage and still scales automatically as traffic grows. Switching models or external providers changes capability rather than fixing the billing model.

Correct answer: A. Evaluate the quality and accuracy of the retrieved context to ensure the right documents are fetched.

Unrelated products in RAG answers usually signal a retrieval problem, so first evaluate whether the right context is being fetched. Swapping the LLM targets generation, which is not the failing stage here.

Correct answer: A. Add a preprocessing step that redacts sensitive information from the Delta table before ingestion.

Redacting sensitive information in a preprocessing step before ingestion ensures the text never reaches the index. Runtime governance filtering still lets sensitive text into the pipeline, so D is weaker when preservation isn't required.

Correct answer: B. Fine-tune the model on internal documents, assuming it will learn what is safe to share.

Fine-tuning on internal documents does not guarantee the model learns access rules, so it can still hallucinate or leak confidential content. Guardrails, role-based access, and careful system prompts are all effective controls.

Correct answer: D. Curate and manually review upstream data before adding it to the RAG pipeline.

Curating and manually reviewing upstream data before ingestion keeps offensive content out of the vector store and therefore out of responses. Restricting who can read the source data doesn't change what the model generates.

Correct answer: A. Provide clear examples in the prompt that demonstrate the expected response format.

Providing clear examples, essentially few-shot prompting, is the best way to steer the model toward a consistent output format. Vague or missing format instructions leave the structure to chance.

Correct answer: A. Ensure each row indexed into the Delta Sync Index does not exceed the documented 100 KB row-size limit.

The Delta Sync Index enforces a documented 100 KB per-row size limit, so keeping each indexed row under it is correct. The other options cite limits that don't match the documented Vector Search constraints, so eliminate them on sight.

Correct answer: E. All of the above.

A well-structured prompt combines all four components: instruction, context, input, and desired output format. Since each individual option is a valid component, the all-of-the-above choice is the only complete answer here.

Correct answer: A. Precision @ K and Faithfulness.

Precision at K measures relevance of the top retrieved chunks, and Faithfulness measures how closely the generated answer sticks to them. Latency and toxicity don't address groundedness, so they don't fit the stated requirement.

Correct answer: B. Apply a toxicity classifier and a PII-detection algorithm to the dataset.

A toxicity classifier removes harmful language and PII detection removes sensitive information, directly meeting both stated goals. Deduplication and chunking improve quality but don't filter harmful or sensitive content.

Correct answer: B. Use the columns_to_sync parameter to specify the primary key and embedding column.

The columns_to_sync parameter explicitly restricts which columns the Vector Search index synchronizes. Sync modes and embedding options control refresh behavior, not column selection, so they can't filter what gets indexed.

Correct answer: A. Encouraging users to submit long, detailed questions instead of concise ones.

The correct answer works because concise queries produce focused embeddings and better retrieval, while verbose questions add noise and can degrade semantic search. For the exam, remember that an embedding model, vector database, and response LLM are all core RAG components.

Correct answer: D. Build an agent with tools for SQL queries on Delta tables and real-time web search, feeding both into the LLM.

An agent with a SQL tool for Delta tables and a web-search tool dynamically combines live structured and unstructured data at query time. Options that pre-summarize or push everything into a vector store fail because they cannot handle real-time, query-dependent lookups.

Correct answer: D. Log the model with MLflow during training, register it directly to Unity Catalog via the MLflow API, and start a serving endpoint.

MLflow logging, Unity Catalog registration, and Model Serving is the simplest governed deployment path on Databricks. The Docker and Flask options require you to build and manage your own infrastructure, which defeats the low-setup goal.

Correct answer: A. Define a system prompt listing the available tools and build an agent that dynamically calls the right tool per query.

A tool-using agent lets the LLM decide per query whether to call the schedule API, run a SQL query, or answer from history. Cramming data into the system prompt or relying on RAG alone cannot handle dynamic, external data sources.

Correct answer: A. Implement a human-reviewed data-curation process to filter and clean source documents before ingestion.

Cleaning and curating source documents before ingestion removes the offensive content that the RAG pipeline would otherwise retrieve and echo. A disclaimer or faster re-indexing does not fix the underlying data-quality problem.

Correct answer: D. Include few-shot examples in the prompt showing the exact summary format without explanations.

Few-shot examples that show the exact desired summary format teach the model to skip self-referential commentary. Chunking or ingestion fixes address retrieval quality, which is not the issue here since the summaries are already good.

Correct answer: C. Encouraging users to ask long, overly detailed questions instead of concise ones.

Verbose user questions add noise, ambiguity, and token overhead, which hurts retrieval accuracy and latency. The embedding model, generating LLM, and vector database are all essential parts of a document-grounded support chatbot.

Correct answer: D. User Query -> Embedding Model -> Vector Search -> Context-Augmented Prompt -> Response-Generating LLM -> Final Response

The correct RAG order embeds the query first, retrieves similar passages via vector search, merges them into a context-augmented prompt, and then generates the final response. Any flow that calls the LLM before retrieval breaks the grounding step.

Correct answer: A. User Query -> Embedding Model -> Vector Search -> Reranker -> Context-Augmented Prompt -> LLM -> Final Response

A reranker sits after vector search retrieves candidate passages but before the context-augmented prompt is assembled. Placing the reranker before embedding or search makes no sense because there is nothing to rerank yet.

Correct answer: C. Apply metadata or corpus filtering during retrieval to retrieve only TechNova-related documents, and reinforce this with a system prompt restricting responses to TechNova content.

Filtering retrieval by metadata or corpus plus a system prompt keeps only relevant documents reaching the model, which is the recommended defense in depth. Prompt-only instruction fails because irrelevant documents still enter the context and can sway the answer.

Correct answer: C. Build an agent with tools for SQL queries on Delta tables and web search, then pass results to the LLM.

An agent with a SQL tool for Delta tables and a web search tool for live headlines handles both structured and real-time data. Pure RAG fails here because live news cannot be pre-embedded into a vector store.

Correct answer: C. Chunk the HR documents, embed them into a vector store, retrieve top-matching chunks for the query, and pass them to the LLM for answer generation.

Chunking, embedding into a vector store, retrieving top matches, and grounding the LLM is the standard RAG design for document Q and A. Averaging whole-document embeddings loses fine-grained signal and hurts retrieval precision.

Correct answer: A. Use Llama Guard through the Foundation Model API to automatically detect and filter unsafe responses.

Llama Guard via the Foundation Model API provides built-in safety filtering with the least setup and maintenance. Regex rules miss most unsafe phrasing, and self-evaluation LLM calls add cost, latency, and complexity.

Correct answer: B. Log the model in Dev with MLflow, register it in the Unity Catalog Model Registry, validate the same registered model version in Test, grant Production access using Unity Catalog permissions, and deploy the same approved model version in Production.

MLflow logging plus Unity Catalog Model Registry promotes one validated model version across environments with governance and lineage intact. Manual file copies break reproducibility, and retraining in Production is wasteful and error prone.

Correct answer: C. Trigger a safety/contextual guardrail that detects the emergency, halts normal generation, and returns an emergency message to contact a doctor or hotline immediately.

A safety guardrail should detect emergency intent, stop normal generation, and return a fixed referral message to a doctor or hotline. Generating advice with a disclaimer is insufficient for potentially serious medical situations.

Correct answer: B. Option 2 – Multimodal embedding (CLIP): more complex and compute-heavy, but provides a unified image-text embedding space, improving retrieval relevance.

A multimodal embedding model like CLIP maps images and text into one shared vector space, enabling true cross-modal semantic retrieval of visuals alongside text. OCR with captions is fast but only matches on extracted text, missing semantic links between diagrams and their content.

Correct answer: B. Number of customer queries successfully handled per unit of time (throughput).

Throughput measures how many queries the assistant handles per unit of time, which reflects real-time production performance and scalability. Leaderboard rankings and training-time perplexity don't measure live serving behavior.

Correct answer: D. The accuracy and relevance of the translated responses.

Accuracy and relevance of the translations are the priority because they determine whether meaning is preserved and the output is trustworthy. Mirroring source-language structure can actually produce poor translations, so B fails.

Correct answer: B. Use a text-classification model to identify and flag problematic content for review.

A text-classification moderation model catches nuanced problematic content that rule-based matching misses and routes it for human review. Regular expressions only catch literal phrases and fail on paraphrased violations, making them weaker here.

Correct answer: C. Multi-agent system

A multi-agent system lets specialized agents handle account, order, and logistics tasks under a supervisor, matching the cross-departmental scope. A single agent or deterministic chain can't scale across that diverse a workflow.

Correct answer: B. Compliance Guardrail

Compliance guardrails enforce regulatory and policy rules, including restricting the agent to approved topics and refusing off-topic requests. Safety is the strongest distractor, but it targets harmful or unsafe content, not topic-scope and policy compliance.

Correct answer: C. Evaluate and improve the retrieved context before generation to pass the most relevant data to the LLM.

When a RAG assistant drifts off topic, fix retrieval quality first through better chunking, filtering, and evaluation before generation. Swapping models or algorithms treats symptoms rather than the actual context being passed in.

Correct answer: A. Haystack focuses on retrieval and Q&A pipelines, while LlamaIndex emphasizes indexing, data connectors, and integrating data with LLMs via structured context.

Haystack is pipeline-centric around retrievers and question answering, while LlamaIndex centers on indexing, connectors, and feeding structured context to LLMs. The other options describe capabilities neither framework actually has.

Correct answer: D. Provisioned throughput embeddings – pre-allocate capacity for bulk embedding and reuse the endpoint for future tickets.

For over a million tickets with thousands of daily embeddings, provisioned throughput gives predictable capacity and lower cost per embedding. Pay-as-you-go suits one-time or sporadic workloads, not sustained high-volume embedding.

Correct answer: C. Encouraging users to submit long, wordy questions.

Encouraging verbose questions adds noise, cost, and latency without improving retrieval quality. The embedding model, vector database, and generation LLM are all core RAG components that directly drive relevant answers.