Databricks Certified Generative AI Engineer Associate Practice Test – 127 Free Exam Questions with Answers

Correct answer: D. D. Provide few-shot examples to the model or adjust the user prompt to guide the system toward the desired output format.

Providing few-shot examples or adjusting the user prompt directly shapes the output format and removes unwanted meta-text. Modifying chunk sizes or ingestion pipelines serves retrieval relevance, which fails to address the generation behavior itself.

Correct answer: B. B. Fine-tune the model on your data, hoping it will automatically learn to avoid inappropriate outputs.

Simply fine-tuning a model on internal data does not guarantee it will prevent hallucinations or automatically enforce data access restrictions. Guardrails, strict system prompts, and explicit permission-based data retrieval are required for security.

Correct answer: D. D. You will be given a transcript of a customer call where the customer asks about product availability. Respond with "In Stock" if the product is available or "Out of Stock" if it's unavailable.

The correct prompt clearly defines the decision rule based on availability and constrains the output to the required labels. The strongest distractor adds unnecessary JSON formatting and fields, which introduces complexity and potential hallucination errors.

Correct answer: A. A. The system should include a prompt, a retriever, and an LLM. The retriever's output is inserted into the prompt, which is then passed to the LLM to generate answers.

A Retrieval-Augmented Generation pipeline dynamically accesses newly published documents, avoiding the high costs of continuous model retraining. Relying solely on prompt engineering introduces stale knowledge, while utilizing agents introduces unnecessary architectural complexity.

Correct answer: B. B. Create a system prompt for the agent listing the available tools, and implement an agent system that runs different calls to handle the queries.

Defining available tools within the system prompt enables the agent to dynamically interact with external APIs and structured tables. Static architectures like basic Retrieval-Augmented Generation lack the operational flexibility required to trigger real-time external lookups.

Correct answer: C. C. Unstructured

The Unstructured library is purpose-built to extract and clean text from complex file formats like PDFs using minimal code. Standard libraries like BeautifulSoup only parse HTML or XML, making them unsuitable for raw document extraction. Rely on Unstructured for efficient data preparation.

Correct answer: C. C. The chatbot should be designed as a multi-step LLM workflow. First, it should identify the type of question being asked, then route the query to the appropriate model. For questions about upcoming events, the query should be directed to a text-to-SQL model, while ticket purchasing inquiries should redirect the user to a payment platform.

Building a multi-step workflow with an initial intent classification step allows the system to route queries to specialized tools or models. Handling everything in one prompt or splitting into multiple chatbots creates unnecessary complexity. Look for options that describe query routing or agentic workflows.

Correct answer: C. C. "Given the complaint, respond in the following JSON format: { "issue_summary": …, "suggested_action": …, "urgency_level": … }."

Providing an explicit JSON schema directly in the prompt forces the model to adhere to the required structure and field names. Simple conversational prompts fail to enforce machine-readable formats reliably. Always use clear formatting instructions when structured outputs are needed.

Correct answer: C. C. Input: production logs; Output: failure reason and recommended fix

Defining inputs as production logs and outputs as generated diagnoses directly aligns with the stated application requirements. Using historical tickets or database tables reverses the data flow. For exam questions mapping pipelines, strictly match the data sources and desired business outcomes.

Correct answer: C. C. Pre-process documents to remove disclaimers, footers, and legal sections before chunking.

Pre-processing documents to remove irrelevant noise before chunking ensures the vector database contains only meaningful context. Increasing chunk overlap or expanding context windows simply adds more noise to the prompt. Cleaning raw data is always the first step for reliable retrieval.

Correct answer: C. C. Use Spark DataFrame API to write as Delta → Save to managed Unity Catalog volume → Register table in appropriate schema

Using Spark to write chunked data into a managed Unity Catalog Delta table provides ACID compliance and centralized security. Saving files as CSV or Parquet bypasses governance features and lacks time travel. Rely on managed Delta tables for secure RAG infrastructure.

Correct answer: A. A. Use evaluation metrics like Mean Reciprocal Rank (MRR) or NDCG to assess ranking quality · E. E. Build a test set of query-answer pairs and compute precision/recall across retrieval strategies

Building a labeled test set and calculating standard metrics like precision provides an objective baseline for comparing retrieval strategies. Rank-aware metrics like Mean Reciprocal Rank measure how well the system orders relevant results. Avoid subjective or isolated component metrics for holistic evaluation.

Correct answer: B. B. Inject the metadata fields into the prompt as system context before the user's input

Injecting metadata directly into the system context grounds the model in the specific user profile before it generates a response. Appending metadata as a footnote happens after generation, meaning it cannot influence the highly personalized output you need.

Correct answer: B. B. Modify the prompt to instruct: "Respond in one sentence. Be direct and actionable."

Explicit prompt instructions directly steer the style and length of the output without requiring any changes to the underlying model. Raising temperature increases randomness and often makes responses longer, while hard token limits risk truncating important information mid-sentence.

Correct answer: B. B. A summarization-capable model with support for structured output

A summarization-capable model with structured output support reliably extracts fields like issue type directly from text transcripts. A speech-to-text model is the wrong choice because the audio transcription step is already completed.

Correct answer: C. C. Register the model to Unity Catalog using MLflow → Apply access controls → Deploy to Model Serving

Registering models in Unity Catalog using MLflow provides the necessary governance, versioning, and lineage tracking before deployment. Saving artifacts directly to DBFS circumvents these centralized security controls and lacks robust audibility.

Correct answer: A. A. Write a custom pyfunc model with pre- and post-processing logic and deploy via Model Serving

A custom pyfunc model packages the entire workflow, including pre-processing and post-processing, into a single, deployable artifact. Splitting logic into multiple scripts hosted externally removes the built-in dependency capture, versioning, and governance essential for production.

Correct answer: B. B. Replace sensitive fields with consistent pseudo-identifiers before chunking to maintain referential meaning.

Replacing sensitive fields with consistent pseudo-identifiers preserves referential integrity while hiding raw values. Pure deletion breaks context and forces hallucinations, whereas metadata filters cannot redact exposed PII within retrieved text.

Correct answer: B. B. Use MLflow to compare prompt variants and log evaluation scores; monitor inference tables for drift and retrieval accuracy.

Using MLflow to compare prompt variants and monitoring inference tables effectively tracks evaluation metrics and data drift. Token counts lack relevance context, and blind fine-tuning without logging hides pipeline failures.

Correct answer: C. C. Use quantitative evaluation metrics such as BLEU, ROUGE, and exact match against ground truth; combine this with latency and cost metrics.

Combining quantitative metrics against ground truth with latency and cost provides the necessary accuracy and budget balance. Subjective ratings or grammar checks ignore regulatory correctness, while choosing the cheapest model sacrifices quality.

Correct answer: D. D. LangChain

LangChain is specifically designed to orchestrate multi-step workflows and chain interactions with large language models. Pandas, TensorFlow, and PySpark handle data processing or model training rather than runtime LLM application orchestration.

Correct answer: D. D. Contact the data curators directly after you've already started using the trained model to inform them.

Informing data curators after already deploying a model violates proactive compliance and intellectual property guidelines. The strongest distractor, contacting them beforehand, fails because it correctly describes a necessary preventative measure.

Correct answer: B. B. Diversity of responses

Response diversity ensures chatbot interactions remain dynamic and engaging, which is critical for retaining users in gaming environments. Randomness lacks contextual control, while repetition creates a stale and predictable conversational experience.

Correct answer: D. D. CodeLlama-34B

CodeLlama is explicitly fine-tuned for code generation and understanding across multiple programming languages. General models like Llama or MPT lack this specialized training. Expect to choose domain-specific foundation models when the use case demands high accuracy for targeted tasks like coding.

Correct answer: D. D. DBRX

DBRX was designed with a massive context window to handle extensive prompts and long-form retrieval inputs. Older open-source models like Llama 2 have strictly limited context windows. Memorize the context limits of core Databricks models like DBRX to answer these requirements correctly.

Correct answer: B. B. Resume Embedder → Retriever → LLM Summarizer

The standard Retrieval-Augmented Generation pattern requires embedding documents, retrieving relevant context via vector search, and summarizing with an LLM. Skipping the embedding step prevents accurate semantic matching. Memorize the core embedding, retrieval, and generation sequence for exam scenarios.

Correct answer: D. D. Add post-processing logic to exclude any response containing unsupported phrases

Post-processing logic acts as an enforceable guardrail by inspecting the generated draft before it reaches the user. While metaprompts help guide behavior, they are advisory and lack the strict, auditable control needed to filter non-compliant medical advice.

Correct answer: B. B. Choose a model with a balance between longer context window and smaller embedding dimension

Balancing an adequate context window with a smaller embedding dimension controls both storage and inference costs in a RAG pipeline. Always picking the largest available context wastes money, as accuracy can actually drop past certain sequence lengths.

Correct answer: B. B. Model flavor, embedding model, retriever, dependencies, model signature, and input examples

Packaging elements like the retriever and model signature inside an MLflow model ensures reproducibility and tracks lineage across Databricks environments. Basic pipeline components alone do not provide the governance and experiment tracking required for enterprise deployment.

Correct answer: C. C. Exclude all datasets with restrictive licensing and replace them with internally generated content or data with clear commercial use rights.

Excluding restrictively licensed datasets and using content with clear commercial rights eliminates infringement risk. Paraphrasing with an LLM or merely adding disclosures does not cure upstream licensing violations.

Correct answer: B. B. Store chunked text in a Delta table within a Unity Catalog volume, applying appropriate schema and permissions

Storing chunked text in a Delta table within Unity Catalog directly enables centralized governance, column-level lineage tracking, and fine-grained access control. Databricks-specific exam cues favor Delta and Unity Catalog over unmanaged file systems because they integrate seamlessly with downstream vector search.

Correct answer: B. B. Create a chain with: text cleaner → sentiment classifier → toxicity detector → summarizer

Building a sequential chain with dedicated text cleaning, sentiment, toxicity, and summarization components provides the necessary modularity and auditability. Exam cues often favor modular pipelines over single prompts because they allow independent logging and transparent safety guardrails.

Correct answer: C. C. LangChain

LangChain provides the necessary abstractions for building a RAG pipeline, including document loaders, text splitters, and wrappers for external APIs. The other tools focus on distributed training, traditional machine learning, or basic NLP preprocessing rather than end-to-end retrieval chaining.

Correct answer: C. C. Reconfigure the retriever using cosine similarity and ensure metadata filtering is enabled

Reconfiguring the retriever and applying metadata filtering directly improves search precision without requiring a costly system rebuild. However, modifying chunk size is also a standard Databricks tuning step, making the options slightly ambiguous for practical RAG optimization.

Correct answer: H. C. Prompt the LLM with multiple documents at once and let it infer jurisdictional compliance from scratch.

Using a sequential chain of discrete tools provides modularity, testability, and reusable components for each pipeline stage. Summarizing before extracting entities destroys intermediate reasoning steps, and single prompts lack traceability.

Correct answer: G. B. Input: Raw document text; Output: Structured JSON with extracted clauses, missing clause types, and recommended additions

Mapping raw text to a structured JSON object cleanly supports extraction, gap analysis, and recommendations. Plain English summaries lose structured clause data, and starting from metadata skips necessary extraction steps.

Correct answer: C. B. Chunk based on document headings and sections to preserve semantic boundaries, then apply re-ranking after retrieval

Chunking by headings preserves semantic boundaries, and adding a re-ranking step optimizes relevance. Fixed-length windows add latency, and discarding short sections risks losing critical risk factor data.

Correct answer: B. B. Add a metaprompt instructing the model: "Only generate code that uses standard Python libraries and has been tested"

Adding explicit metaprompt instructions constrains the model to standard libraries, directly mitigating package hallucinations while preserving overall flexibility. Avoid extreme temperature limits or rigid deterministic rules because they destroy the creative problem-solving utility expected from large language models.

Correct answer: D. C. Format the prompt to say: "Respond with a JSON object using fields: summary, action, confidence_score. Do not use natural language outside the JSON."

Formatting the prompt to explicitly request a JSON object with specific fields reliably enforces machine-readable outputs. Temperature zero reduces randomness but does not guarantee structure, whereas few-shot examples merely nudge style.

Correct answer: E. D. Prompt the LLM with raw logs and ask it to identify and escalate the most severe events in one response

Using modular tools provides structured, traceable multi-stage reasoning with clear evaluation hooks. Single prompts lack traceability, and manual escalation fails to meet the need for automated, multi-criteria evaluation.

Correct answer: B. A. Use pytesseract to extract text, followed by post-processing to clean noise

Using an established OCR library minimizes development time and integrates easily with Python pipelines. HTML parsers cannot read image pixels, and training a custom model introduces unnecessary complexity and cost.

Correct answer: B. A. Apply dense passage retrieval (DPR) with shorter chunk lengths and compare Recall@k · D. C. Use a test set of known queries and expected chunks to compute NDCG and precision

The correct actions use dense passage retrieval with shorter chunks and validate improvements using labeled test sets. For the exam, pair structural changes like chunk size with objective offline metrics like NDCG rather than subjective checks.

Correct answer: B. B. Choose a smaller language model that is specifically trained for the cancer research domain.

Choosing a smaller, domain-specific language model optimizes the balance of inference cost and task-specific accuracy. For the exam, remember that restricting document access or user queries harms the application's utility and does not scale appropriately.

Correct answer: C. C. Select an appropriate evaluation metric (e.g., recall or NDCG) and experiment with variations in the chunking strategy, such as splitting by paragraphs or chapters, to identify the best-performing approach. · E. E. Develop a metric where the LLM acts as a judge, scoring how well previous questions are answered by the retrieved chunks. Adjust the chunking parameters based on the results of this metric.

Options C and E are correct because systematically optimizing chunking requires defining measurable retrieval metrics or using an LLM-as-a-judge to evaluate ground truth. Simply asking the model to guess token counts or changing unrelated pipeline components does not measure actual retrieval performance. Expect questions on evaluating chunking variations.

Correct answer: B. B. Ingest documents from a source → Index the documents and store them in Vector Search → Develop and test the RAG chain with representative queries → Retrieve relevant documents and generate responses using the LLM → Evaluate the RAG application → Deploy the application using Model Serving

Option B is correct because building a RAG application on Databricks follows a specific workflow of ingesting data, indexing it, testing the chain, evaluating performance, and finally deploying. You must always evaluate the retrieval and generation quality before pushing the application to production. Look for the standard sequence.

Correct answer: C. C. Feature Serving

Option C is correct because Databricks Feature Serving provides low-latency, real-time access to structured features like live sports scores. Foundation Model APIs generate text, but they cannot fetch live external data independently. Rely on Feature Serving to bridge real-time data streams with deployed models.

Correct answer: D. D. Inference Tables

Option D is correct because Databricks Inference Tables automatically capture incoming requests and outgoing responses for Model Serving endpoints. This native feature eliminates the operational overhead of building custom logging microservices. Rely on Inference Tables for built-in production monitoring and auditing.

Correct answer: D. D. Break the HR documentation into chunks and store them in a vector database. Use the employee's question to retrieve the most relevant chunks, and use the LLM to generate a response based on the retrieved documentation.

Option D is correct because breaking documentation into chunks and storing them in a vector database enables targeted, cost-effective retrieval before generation. Averaging embeddings or summarizing entire documents loses the granular context required for accurate answers. Stick to standard retrieval-augmented generation patterns for document tasks.

Correct answer: B. B. Create a dataset to separately test the retrieval and generation components of the system. Utilize MLflow's built-in evaluation metrics for this assessment.

Evaluating retrieval and generation components separately using MLflow metrics pinpoints specific failure areas in a RAG pipeline. Option A fails because cosine similarity alone cannot assess generation fluency or correctness, making a holistic, component-level approach necessary for optimization.

Correct answer: B. B. Log the model using MLflow during the training phase, register the model directly in Unity Catalog via the MLflow API, and start a serving endpoint.

Logging the model with MLflow, registering it in Unity Catalog, and starting a serving endpoint is the most seamless deployment path on Databricks. Option A fails because pickling models introduces security risks and bypasses the native governance provided by MLflow model tracking.

Correct answer: B. B. Deploy the model using pay-per-token throughput, as it provides cost guarantees.

Pay-per-token throughput provides cost effectiveness for applications with unpredictable or low traffic volumes by aligning expenses directly with usage. Provisioned throughput guarantees capacity but wastes resources for low-volume applications, making it an inefficient strategy here.

Correct answer: B. B. Utilize a neutralizer to standardize the tone and style of the source documents.

Utilizing a neutralizer strips stylistic variations from source documents, directly reducing the linguistic richness needed to generate haikus. The other options actively steer the model through direct instructions, few-shot examples, or fine-tuning.

Correct answer: D. D. Create an agent equipped with tools for SQL querying of Delta tables and web searching, and provide the retrieved data to an LLM for response generation.

Creating an agent with SQL and web search tools enables the dynamic retrieval of real-time stock prices and current news. Option C fails because pre-storing news in a vector database leads to stale information, whereas agents fetch live data during execution.

Correct answer: B. B. Use LangChain's Agent Framework with distinct tools exposed to agents for parsing, interpreting, and suggesting

Using an agent framework with distinct tools supports asynchronous execution and multi-stage reasoning for complex workflows. Option A fails because chaining tasks in a single prompt hides intermediate steps and prevents dynamic function exposure.

Correct answer: B. B. Evaluate responses using an LLM-as-a-judge setup that scores correctness, completeness, and alignment with legal standards

Using an LLM-as-a-judge setup to score responses provides an automated, repeatable way to catch misrepresentations before production. Option D fails because temperature only controls variation; deterministic settings do not inherently fix hallucinations.

Correct answer: B. B. Rewrite the prompt as a metaprompt that instructs the LLM to retain all technical terms and produce a structured summary with sections like "Background", "Methods", and "Findings"

Rewriting the prompt as a detailed metaprompt directly forces the model to preserve terminology and adhere to a specific structure. Option D fails because relying on zero-shot prompting with appended examples lacks the strict constraints needed for formatting.

Correct answer: D. D. Summarization

Summarization is the correct NLP task because it specifically condenses longer text into a shorter form while preserving core information. The other options are incorrect because text classification assigns labels, and a sentencizer merely splits text without summarizing.

Correct answer: A. A. Context length 512: smallest model size is 0.13 GB with embedding dimension 384

Choosing the 512 context length model with the smallest footprint keeps latency and compute costs minimized for chunked retrieval. The other options waste memory and compute by providing unnecessary context limits and larger embedding dimensions for the given chunk size.

Correct answer: D. D. Create a tool that finds available team members for the given project dates. Embed team member profiles in a vector store and use the project description to search and filter for the best-matched team members who are available.

Option D is correct because embedding employee profiles into a vector store and filtering by availability combines structured metadata filtering with unstructured semantic search. This approach efficiently scales compared to keyword matching or brute-force similarity scoring. Remember that Vector Search natively handles hybrid queries.

Correct answer: D. D. Carefully curate the upstream data, including performing manual reviews before incorporating it into the RAG system.

Option D is correct because carefully curating upstream data prevents offensive or inflammatory content from ever entering the retrieval pipeline. Simply refreshing data or notifying users does not actively reduce harmful outputs at the source. Remember that strong guardrails rely heavily on data quality and curation.

Correct answer: D. D. Context length of 512; the smallest model size is 0.13GB, with an embedding dimension of 384.

Option D is correct because choosing the smallest model size and matching context length directly minimizes compute cost and latency. Since the application prioritizes speed over quality, larger models and context windows introduce unnecessary computational overhead. Always align resource selection with strict application constraints.

Correct answer: A. A. Evaluate the quality of the context being retrieved.

Option A is correct because evaluating retrieved context isolates the root cause of irrelevant responses before modifying the language model. If the system retrieves the wrong documents, even the best model will fail to answer correctly. Always debug the retrieval step first when facing relevance issues.

Correct answer: D. D. Create an MLflow PyFunc model that includes a separate function for processing the prompts.

Creating an MLflow PyFunc model is the correct choice because it modularizes and wraps your custom preprocessing logic seamlessly for model serving. For the exam, avoid options suggesting internal model architecture changes, as large language models are treated as fixed inference endpoints.

Correct answer: B. B. Enhance the computing resources to boost the processing speed of questions, allowing for better relevance analysis.

Enhancing computing resources is the correct answer because increasing processing speed or throughput does not inherently improve the topical relevance of the generated outputs. Relevance is achieved through domain guardrails and manual reviews, whereas compute simply makes existing pipelines run faster.

Correct answer: B. B. Flatten the dataframe so that each chunk is in its own row, create a unique identifier for each row, and save it to a Delta table.

Flattening the dataframe so each chunk is an independent row is correct because vector search indexes require granular text segments to perform accurate similarity searches. Avoid storing data as independent JSON files, as Delta tables provide the necessary querying, indexing, and scalability foundation.

Correct answer: B. B. Use a universal multilingual embedding model and apply metadata-based filtering during retrieval to select language-matched chunks

A universal multilingual embedding model maps different languages into a shared space, while metadata filters restrict retrieval to the correct language. For the exam, remember that engine-level metadata filtering in Vector Search prevents irrelevant chunks from polluting the context window.

Correct answer: B. B. Use Databricks ai_query() within a PySpark job to apply the LLM across rows in parallel and log inference results

Using the ai_query function within a PySpark job applies the LLM across rows in parallel, scaling effortlessly for batch workloads. Synchronous REST calls via Model Serving are meant for real-time requests and lack inherent distributed batch parallelism.

Correct answer: B. B. Use masking techniques during chunking to redact PII and enforce license-aware filtering during retrieval

Masking PII during chunking and enforcing license-aware filtering during retrieval guarantees sensitive data never reaches the model. Relying on prompt instructions is a soft guardrail, whereas engine-level metadata filters provide a hard, reliable boundary.

Correct answer: B. B. Apply content filtering during retrieval to exclude toxic language before passing chunks to the model

Applying content filtering during retrieval excludes toxic language before the model receives the context, preserving safe, useful data. Removing all chat logs sacrifices valuable coverage, while relying on the foundation model risks quoting toxic text verbatim.

Correct answer: A. A. Track latency by language using inference tables and set hard limits on total token count per request

Tracking latency by language using inference tables and setting hard limits on total token count per request addresses both issues. For the exam, remember that Databricks inference tables provide granular telemetry for performance debugging, while max_tokens limits directly control generative costs.

Correct answer: A. A. Increase the document chunk size · B. B. Reduce the overlap between chunks

Increasing the document chunk size and reducing the overlap between chunks both directly reduce the total number of generated vectors. Remember that switching to a smaller embedding model only reduces vector dimensionality and memory footprint, not the actual embedding count.

Correct answer: C. C. pytesseract

The pytesseract library provides a simple Python wrapper for optical character recognition to extract text from image files. The other options handle HTML parsing or web scraping, making them completely ineffective for reading text embedded inside image formats.

Correct answer: C. C. Add credentials through environment variables.

Adding credentials through environment variables is correct because it securely externalizes secrets from the model code during endpoint deployment. As a practical exam cue, remember that hardcoding secrets or using Spark configurations exposes sensitive data, whereas environment variables safely integrate with Databricks secret management.

Correct answer: A. A. Implement a safety filter that identifies harmful inputs and instruct the LLM to inform the user that it cannot assist.

Implementing a safety filter is correct because it actively intercepts and blocks harmful or inappropriate user inputs before the model processes them. As an exam cue, remember that allocating more compute or limiting interaction time does not provide content moderation or protect against malicious prompts.

Correct answer: E. C. Contextual Guardrail

A contextual guardrail is the correct choice because it restricts the large language model to approved business topics, such as insurance, preventing off-topic drift. For the exam, remember that safety guardrails block harmful or toxic content, whereas contextual guardrails enforce business scope and relevance.

Correct answer: C. C. State in the system prompt that the application is not intended to answer questions unrelated to SnoPen AI.

Stating in the system prompt that the application is not intended to answer unrelated topics is correct because it explicitly sets the operational boundaries for the model. As a practical exam cue, remember that combining all documents into a single chunk severely degrades retrieval accuracy and increases token costs.

Correct answer: C. B. The pyfunc model was registered with only the base LLM and not the wrapped chain logic

Registering only the base LLM instead of the wrapped pyfunc object bypasses any custom predict methods. Remember that Databricks Model Serving executes the specific artifact registered, so unserialized preprocessing logic will be completely ignored at runtime.

Correct answer: D. C. Apply metadata filtering during query execution to narrow vector search scope before similarity ranking

Applying metadata filters during the query narrows the vector search scope before similarity ranking, reducing latency. Do not rely on the LLM to filter text, as it requires retrieving irrelevant vectors first, which degrades performance and increases token costs.

Correct answer: H. C. Register the model to Unity Catalog under a shared catalog and enable model serving with access rules for each team

Registering the model in Unity Catalog provides native version control, lineage tracking, and auditable governance. Serving directly from DBFS or an external VM fails to provide the fine-grained access controls required across enterprise teams.

Correct answer: G. A. Deploy the chain using Databricks Model Serving with MLflow; enable inference logging and configure observability with Databricks metrics

Deploying the chain with Databricks Model Serving and MLflow enables scalable real-time REST endpoints for APIs. Enabling inference tables is your key exam cue for automatically tracking traffic, latency, and full user interactions in a governed Delta table.

Correct answer: G. B. Recreate the vector index regularly with updated chunks from source documents

Recreating or syncing the vector index ensures the latest document chunks are embedded and available for retrieval. Prompting the model to ignore old policies fails because the retriever never surfaces the new chunks for the model to see.

Correct answer: E. A. Use NDCG and Recall@k to assess retrieval accuracy across both configs, and complement with human judgment for output quality

Using retrieval metrics like NDCG and Recall provides an objective measure of how well each config surfaces the right context. Relying solely on token usage or chunk quantity ignores accuracy, which is critical for safe legal applications.

Correct answer: D. B. Please call your local emergency services.

Directing the user to call emergency services is correct because severe headaches and dizziness represent an urgent medical triage scenario requiring immediate human care. For the exam, always prioritize safety and escalation over standard information gathering when a user describes potentially severe medical symptoms.

Correct answer: D. D. The accuracy and relevance of the generated responses.. The degree of similarity to the original language.

Assessing the accuracy and relevance of the generated responses is correct because safe translation requires preserving the original meaning without introducing harmful, biased, or misleading context. The distractor focusing on similarity to the original language fails because direct literal translation often misses cultural nuances.

Correct answer: A. What is the BEST strategy to implement this enrichment while preserving model accuracy and performance?

Injecting targeted structured metadata into the prompt supplies the grounding context needed to disambiguate vague queries. Option B fails because appending the entire user history inflates token costs and risks exposing irrelevant data.

Correct answer: B. Register the model directly in Unity Catalog using MLflow, then grant the development workspace READ permissions on that model.

Registering the model in Unity Catalog enables secure, cross-workspace sharing with fine-grained access controls without duplicating data. This approach avoids the storage and compute costs of manual exports or redundant training pipelines, aligning with standard Databricks governance practices.

Correct answer: A. Massive Multi-task Language Understanding (MMLU) benchmark score

Massive Multi-task Language Understanding, or MMLU, is an academic benchmark meant for testing broad foundation model capabilities rather than measuring live application success. For the exam, remember that production systems require operational metrics like accuracy, latency, and throughput, not generic research scores.

Correct answer: C. Retain all articles—including non-technical ones—because the system needs exposure to unrelated content in order to discourage off-topic answers

Retaining irrelevant documents pollutes the vector store and defeats the core purpose of retrieval-augmented generation by increasing noise and hallucination risk. A key exam cue is that curating and filtering the corpus before indexing is a fundamental best practice for domain-specific systems.

Correct answer: C. Build a dedicated evaluation dataset that independently tests retrieval effectiveness and generation quality, and use MLflow's evaluation features to assess each component separately.

Building a dedicated evaluation dataset and using MLflow allows engineers to assess retrieval and generation independently, isolating bottlenecks accurately. Using ROUGE or an LLM-as-a-judge only grades final text, hiding whether failures stem from poor context fetching or weak generation.

Correct answer: A. It enables complex tasks to be divided into a sequence of smaller, dependent steps, allowing the assistant to produce more thorough and accurate answers.

Prompt chaining divides complex tasks into smaller sequential steps, passing outputs from one stage as inputs to the next for improved accuracy. Chaining actually increases latency and token costs, making options claiming speed or cost reduction clearly incorrect.

Correct answer: C. Inference Tables

Inference Tables automatically log requests and responses for provisioned-throughput endpoints, providing the exact observability needed. Be careful not to confuse this with Vector Search, which handles embedding retrieval for RAG rather than tracking endpoint payloads.

Correct answer: B. Explode the chunk array so each row contains a single chunk, generate a unique ID for every row, and enable change data feed on the resulting Delta table.

Exploding the array ensures each chunk is an independent row, generating unique IDs creates a required primary key, and enabling change data feed allows efficient synchronization. Writing unchanged data fails because Vector Search requires a unique primary key for every indexable record.

Correct answer: C. Use a framework such as ReAct, enabling the LLM to create reasoning traces and invoke external tools through structured actions whenever required.

The ReAct framework enables an LLM to interleave reasoning traces with autonomous external tool actions, supporting dynamic planning and revision. Standard Chain-of-Thought prompting improves reasoning but requires manual intervention to supply tool outputs, preventing true autonomous execution.

Correct answer: A. Authenticate using an OAuth flow designed for machine-to-machine communication · B. Use an access token issued to a service principal

Using an access token issued to a service principal provides the non-interactive, machine-to-machine authentication required for secure production deployments. While machine-to-machine OAuth is conceptually similar, explicitly using a service principal ensures independent lifecycle management and strict least privilege.

Correct answer: A. Add the section header as a prefix to chunks · D. Increase the document chunk size

Adding section headers and increasing chunk size directly resolve incomplete retrieval by preserving semantic context within structured documents. Sentence splitting destroys context, and upgrading models cannot fix missing information reaching the response generator.

Correct answer: B. Deploy Llama Guard through the Foundation Model API and use it to screen outputs for safety issues

Deploying Llama Guard through the Foundation Model API provides a purpose-built moderation layer with minimal engineering overhead. Regex rules are too brittle for semantic safety, while custom LLM calls introduce unnecessary cost, latency, and complexity into the application flow.

Correct answer: A. BLEU

BLEU is the industry standard metric for evaluating machine translation quality by measuring n-gram overlap against high-quality reference translations. ROUGE is used for summarization, while NDCG and Recall apply to ranking and retrieval tasks rather than translation fidelity.

Correct answer: C. Use an embedding model that was trained on data similar to the application's domain

Selecting an embedding model trained on similar domain data ensures it accurately captures the specific semantics and vocabulary of your corpus. General leaderboard rankings or multilingual capabilities do not guarantee strong performance on your highly specialized internal documents.

Correct answer: D. beautifulsoup

BeautifulSoup is a Python library built specifically for parsing HTML and XML, extracting text with minimal code. Exam tip: pytesseract is for OCR on images, PyPDF2 is for PDF files, and NumPy handles numerical computing.

Correct answer: D. Collect the policy documents → Create embeddings and store them in Vector Search → User sends a query to the LLM → LLM retrieves relevant documents → LLM generates an answer → Evaluate performance → Deploy using Model Serving

The correct workflow begins with collecting documents, creating embeddings, and storing them in Vector Search before users submit queries. Evaluation must occur before deployment to ensure the pipeline retrieves context accurately and generates grounded responses in production.

Correct answer: D. Apply rate-limiting controls directly on the Foundation Model endpoints.

Applying rate-limiting controls directly to the endpoint enforces hard infrastructure quotas, effectively preventing accidental traffic spikes. Distractors relying on prompt instructions or manual code reviews fail because they lack real-time, automated enforcement at the serving layer.

Correct answer: B. Add metadata fields for publication date and topic so the system can apply metadata filters during retrieval.

Adding structured metadata fields like publication date and topic allows the system to apply hard filters during vector retrieval. This is the most efficient approach because relying on semantic similarity alone cannot guarantee precise constraints like specific date ranges.

Correct answer: D. Deploy the endpoint with provisioned throughput to obtain predictable, guaranteed performance.

Provisioned throughput reserves dedicated compute capacity for a foundation model endpoint, guaranteeing predictable performance and avoiding noisy neighbor issues. For the exam, associate provisioned throughput with reliable production scaling rather than simply shrinking the model size.

Correct answer: A. Break the manuals into smaller text chunks and generate embeddings stored in a vector database. For each customer query, retrieve the most relevant chunks and have the LLM craft an answer using the retrieved content.

Breaking manuals into chunks and retrieving the most relevant sections forms the standard retrieval-augmented generation pipeline, ensuring accurate and efficient answers. Avoid options involving collaborative filtering or massive context windows, which are inefficient or lossy.

Correct answer: D. Add a safety-filtering layer that flags harmful text and instructs the model to decline those requests.

A dedicated safety-filtering layer intercepts and blocks harmful prompts before they reach the language model, preventing malicious inputs from generating unsafe outputs. Relying on the model to warn users leaves the system vulnerable, whereas compute scaling or time limits never address input toxicity.

Correct answer: B. Allow the system to ingest content only from vetted, trustworthy social media and news accounts so unexpected toxic material is excluded.

Curating data sources to ingest content only from trusted accounts prevents toxic material from entering the retrieval corpus, which keeps harmful context from influencing model outputs. Limiting retrieval count fails because any selected toxic chunk still poisons the generation phase.

Correct answer: A. The Big Code Models Leaderboard · B. The model's parameter count

The Big Code Models Leaderboard evaluates task-specific coding capabilities, making it essential for choosing a code generation base model. Parameter count dictates GPU memory requirements and directly impacts hosting costs, whereas MTEB only ranks embedding models.

Correct answer: A. Turn on MLflow tracing to obtain detailed visibility into the agent's execution flow and the behavior of each step.

MLflow tracing provides granular observability into agent execution flows, automatically capturing inputs, outputs, and tool calls for multi-step workflows. Structured logging lacks this deep automated introspection, while MLflow evaluate is meant for scoring model quality.

Correct answer: A. Introduce a reranker to reorder the retrieved documents based on their true relevance scores.

Introducing a reranker resolves this by using a cross-encoder to reorder the retrieved documents based on true semantic relevance. Relying on the LLM to reorder context is unreliable, whereas a dedicated reranking stage is the standard approach to fix imperfect embedding similarity.

Correct answer: D. whisper-large-v3 (1.6B)

Whisper-large-v3 is an open-source model specifically designed for fast and accurate automatic speech recognition. For the exam, remember that audio transcription requires specialized acoustic models like Whisper, whereas text-based models like DBRX cannot process audio natively.

Correct answer: A. Use only datasets that are clearly marked with an open license and carefully adhere to all associated licensing terms.

Using datasets with clear open licenses and strictly adhering to their terms creates a reliable chain of rights, reducing legal exposure. Treating public data as unrestricted is a major trap, as public accessibility does not eliminate copyright or usage constraints.

Correct answer: B. You will receive a customer call transcript involving a question about product availability. Output "Success" if the product is available and "Fail" if the product is unavailable.

Option B provides direct, deterministic instructions that align perfectly with the strict product availability classification task. Avoid options introducing unnecessary reasoning steps or complex formatting, as these distract the model and degrade classification consistency.

Correct answer: A. Vector Stores · B. Conversation Buffer Memory

Vector stores are essential for retrieving relevant knowledge, while conversation buffer memory maintains context across multiple chat turns. External tools and UI components might enhance an application, but they are not foundational requirements for a standard conversational RAG pipeline.

Correct answer: A. CodeLlama-34B-Instruct-hf

CodeLlama is explicitly pretrained and instruction-tuned for code generation, making it the optimal specialized model. General-purpose models like Llama-2 or Mixtral can write code, but they lack the targeted architecture needed for high-quality out-of-the-box programming tasks.

Correct answer: A. Encouraging customers to ask lengthy, detailed questions instead of short ones

Encouraging long questions does not improve a RAG chatbot, which relies on semantic embeddings to understand brief queries naturally. The LLM, embedding model, and vector database are foundational components, making the query length a irrelevant design factor.

Correct answer: C. Craft prompts that explicitly instruct the LLM to produce output in the required tone and style

Crafting explicit prompts is the most efficient initial method to steer model behavior because it adjusts output without modifying weights or requiring new pipelines. For the exam, always choose prompt engineering over fine-tuning when the goal is a low-cost, immediate adjustment to tone or style.

Correct answer: D. Build a multi-step chain where one LLM first identifies sentiment and, based on that classification, dynamically adjusts the system prompt used by the response-generation LLM.

Building a multi-step chain allows a dedicated model to classify sentiment and dynamically adapt the system prompt for the final response. This orchestration pattern is preferred over retrieval or basic regression because LLMs handle nuanced emotional cues and tone shifts much more effectively.

Correct answer: A. Provide few-shot examples that clearly demonstrate the exact output format expected

Providing few-shot examples reliably forces the model to mimic a strict output format without generating unwanted explanations. Zero-shot instructions are often ignored for formatting, and chain-of-thought prompting would actually increase the reasoning text you want to eliminate.

Correct answer: A. Input: Customer service chat histories; Output: Organize chats by user, summarize each user's prior interactions, and then craft a response

Organizing and summarizing chat histories by user enables faster responses while maintaining a highly personalized experience based on prior context. Retrieval pipelines lack personalization, and sentiment analysis or clustering reviews does not directly generate a reply to customer inquiries.

Correct answer: C. Build the chatbot as a multi-stage LLM pipeline: first classify the user's query type, then route it accordingly. Direct failure-rate questions to a text-to-SQL model that queries the failure table, and send troubleshooting queries to another model that summarizes the relevant manual and produces an answer.

A multi-stage pipeline classifies user intent and routes the request to the appropriate downstream handler, such as text-to-SQL for structured data or retrieval for manuals. Combining all data into a single format degrades accuracy by stripping away the native structure of relational tables.

Correct answer: A. Use windowed chunking with overlapping segments to maintain continuity across adjacent text.

Windowed chunking with overlapping segments maintains continuity across adjacent text boundaries, ensuring legal reasoning remains intact. Paragraph chunking lacks overlap causing context gaps, while sentence chunking is too granular, and hierarchical clustering disrupts chronological flow.

Correct answer: B. 1. Define or import the tools 2. Specify the agent configuration 3. Initialize the agent with ReAct logic, the LLM, and the tools

In LangGraph, the correct sequence is to define the tools, specify the agent configuration, and then initialize the agent with ReAct logic. Unlike classic LangChain, LangGraph requires defining the graph state and transitions before instantiating the actual agent.

Correct answer: A. The OpenAI client library

The OpenAI client library requires its own explicit API key and does not inherit Databricks notebook session credentials. For the exam, remember that native tools like the Databricks Python SDK and MLflow Deployments automatically authenticate using the active workspace session.

Correct answer: A. Measure the cosine similarity between the embeddings of retrieved items and those of a representative set of test queries

Cosine similarity measures the semantic closeness between query and retrieved item embeddings, directly evaluating retrieval accuracy. Avoid text generation metrics like BLEU or ROUGE, which measure n-gram overlap for translation or summarization tasks instead of vector distance.

Correct answer: B. Use Databricks access control lists (ACLs) to tightly manage which identities can interact with the model.

Applying access control lists directly restricts which identities can query the model endpoint, mitigating inversion risks. While guardrails are useful, they filter unsafe content rather than stopping adversarial probing, making strict endpoint access controls the strongest defense.

Correct answer: D. Review the model card and training data details to verify whether the model was trained on medical or biomedical corporathe model with the highest download count, assuming that popularity reflects broad effectiveness and model quality

Reviewing the model card and training data details confirms whether the model was exposed to biomedical corpora, which is critical for domain-specific accuracy. For the exam, always rely on documented training data over proxy metrics like popularity or recency when selecting models.

Correct answer: F. Using extremely large chunk sizes that stretch to the limits of the model's context window in hopes of maximizing text coverage

Using extremely large chunks degrades retrieval precision by packing unrelated information together, which increases noise and hallucinations. Metadata filtering, few-shot examples, and custom embedding fine-tuning are all effective strategies that meaningfully improve chatbot accuracy.