View Full Databricks Certified Generative AI Engineer Associate Exam Dumps and Practice Test Dumps

 

Question 81. Which technique can help improve the quality of retrieved documents in a RAG application?

1. Increasing the model temperature
   2. Improving the embedding or retrieval strategy
   3. Removing all metadata
   4. Increasing response length

Correct Answer: 2. Improving the embedding or retrieval strategy

Explanation:

Improving the embedding or retrieval strategy can significantly increase the quality of context supplied to a generative AI model in a RAG application. Embeddings should represent the meaning of documents and user queries effectively so that semantically relevant information can be retrieved. Retrieval quality can also be improved through appropriate chunking, metadata filtering, query rewriting, hybrid search, or reranking. Increasing the model temperature does not improve document retrieval because temperature controls generation behavior rather than search quality. Similarly, removing metadata can make filtering and contextual relevance worse. Better retrieval ensures that the generation model receives useful, accurate, and relevant information, which can improve grounded responses and reduce hallucinations.

Question 82. What is the primary purpose of reranking in a retrieval-augmented generation pipeline?

1. Generate the final response
   2. Create embeddings for documents
   3. Reorder retrieved results according to relevance
   4. Store documents in a vector database

Correct Answer: 3. Reorder retrieved results according to relevance

Explanation:

Reranking is used after an initial retrieval stage to improve the ordering of candidate documents according to their relevance to the user’s query. A retrieval system may return several potentially useful documents, but the initial similarity search does not always place the most useful passage first. A reranker evaluates the query and retrieved candidates using a more detailed relevance assessment and assigns a new ordering. This allows the generation model to receive the strongest evidence near the top of its context. Reranking does not create embeddings, store documents, or generate the final answer. It is primarily a retrieval-quality improvement technique.

Question 83. Which practice is most appropriate for protecting sensitive information in a generative AI application?

1. Apply access controls and minimize sensitive data exposure
   2. Include all available data in every prompt
   3. Store credentials directly in prompts
   4. Disable authentication for internal users

Correct Answer: 1. Apply access controls and minimize sensitive data exposure

Explanation:

Protecting sensitive information requires controlling who can access data and ensuring that applications expose only the information necessary for a particular task. Access controls can restrict users and services according to their authorized permissions, while data minimization reduces unnecessary exposure of confidential information. Sensitive values should not be placed directly into prompts unless they are required and appropriately protected. Credentials should never be embedded in application prompts or source code. Authentication and authorization should remain enabled even when an application is considered internal. These practices reduce the risk of unauthorized disclosure and help establish safer generative AI workflows.

Question 84. Why is chunk overlap sometimes used when preparing documents for RAG?

1. To increase model temperature
   2. To eliminate embeddings
   3. To prevent the need for retrieval
   4. To preserve context across chunk boundaries

Correct Answer: 4. To preserve context across chunk boundaries

Explanation:

Chunk overlap helps preserve information that might otherwise be split between two separate document chunks. When a document is divided into independent segments, an important sentence or concept can occur near the boundary between chunks. Without overlap, one chunk may contain the beginning of the idea while the next contains the remaining details. Adding a controlled amount of overlap allows related content to appear in more than one chunk, increasing the chance that retrieval returns enough context to answer a query correctly. However, excessive overlap can increase storage requirements, retrieval redundancy, and the amount of duplicated context provided to the model.

Question 85. What does a context window determine for a large language model?

1. The number of database records it can permanently store
   2. The amount of input and output tokens it can process within a request
   3. The number of users allowed to access an endpoint
   4. The number of vector indexes that can be created

Correct Answer: 2. The amount of input and output tokens it can process within a request

Explanation:

A model’s context window defines how much tokenized information it can consider within a particular interaction or request. This generally includes relevant portions of the user prompt, system instructions, retrieved documents, and conversation history, along with the generated response subject to the model’s limits. The context window is not a measure of permanent database storage or the number of users an endpoint can support. In RAG applications, context-window size is especially important because retrieving too much information can exceed the model’s limits or reduce the useful signal available to the model. Effective chunking and retrieval help keep supplied context focused and manageable.

Question 86. Which evaluation metric is most directly concerned with whether an answer is supported by retrieved information?

1. Groundedness
   2. Throughput
   3. Endpoint availability
   4. Token generation speed

Correct Answer: 1. Groundedness

Explanation:

Groundedness evaluates whether a generated response is supported by the information available to the application, such as retrieved documents or supplied source context. This is particularly important for RAG systems because the application is expected to use authoritative information rather than invent unsupported facts. A highly grounded response should make claims that can be traced back to the provided evidence. Metrics such as throughput and token generation speed measure system performance rather than factual support. Availability measures whether a service is operational. Groundedness therefore provides a useful quality signal when evaluating whether a generative AI application is using retrieved evidence appropriately.

Question 87. What is a major advantage of metadata filtering in a vector search system?

1. It guarantees that every generated answer is correct
   2. It eliminates the need for embeddings
   3. It restricts retrieval to documents matching specified attributes
   4. It automatically fine-tunes the language model

Correct Answer: 3. It restricts retrieval to documents matching specified attributes

Explanation:

Metadata filtering allows a retrieval system to limit candidate documents according to attributes associated with those documents. For example, an application could retrieve only documents belonging to a particular department, product, language, date range, or access category. Combining metadata filters with semantic similarity can improve relevance and reduce the likelihood of retrieving information from inappropriate sources. Metadata filtering does not guarantee that an answer is correct, nor does it eliminate the need for embeddings when semantic retrieval is being used. It also does not fine-tune the language model. Instead, it provides an additional mechanism for controlling which records are eligible for retrieval.

Question 88. Which approach is useful when users ask questions using terminology different from the wording in the source documents?

1. Disable retrieval
   2. Query rewriting or expansion
   3. Remove document embeddings
   4. Increase database storage

Correct Answer: 2. Query rewriting or expansion

Explanation:

Query rewriting or expansion can help when the terminology used by a user differs significantly from the language contained in the source documents. A rewriting process can transform the original question into a clearer or more retrieval-friendly query, while expansion can introduce relevant synonyms, related terms, or alternative formulations. This can increase the probability that semantic or lexical retrieval finds useful documents. Simply disabling retrieval would remove the application’s ability to obtain external evidence, while increasing storage does not directly address query-document mismatch. Query transformation is therefore a practical technique for improving retrieval performance when users and source material use different terminology.

Question 89. Why should prompt templates be version controlled?

1. To track changes and compare their effects on application behavior
   2. To permanently increase model context windows
   3. To replace authentication
   4. To prevent all model errors automatically

Correct Answer: 1. To track changes and compare their effects on application behavior

Explanation:

Version controlling prompt templates makes it possible to track changes made during application development and understand how those changes affect model behavior. A prompt may influence retrieval, reasoning, formatting, safety behavior, and response quality, so changing it without recording the modification can make evaluation and troubleshooting difficult. By maintaining versions, developers can compare experiments, reproduce previous results, identify which prompt revision introduced a regression, and roll back to a known configuration when necessary. Prompt version control does not eliminate model errors or replace authentication. It is primarily a software-development and evaluation practice that supports reproducibility and controlled iteration.

Question 90. What is the purpose of an evaluation dataset for a generative AI application?

1. To permanently store model weights
   2. To replace the production database
   3. To increase endpoint network bandwidth
   4. To provide representative examples for measuring application quality

Correct Answer: 4. To provide representative examples for measuring application quality

Explanation:

An evaluation dataset contains representative inputs, expected behaviors, reference answers, or other criteria that can be used to measure the quality of a generative AI application. It allows developers to test an application consistently as prompts, retrieval strategies, models, and other components change. A useful evaluation dataset should represent realistic user questions and important edge cases rather than containing only easy examples. Comparing results against the dataset can reveal regressions and help teams determine whether a modification improves relevant quality measures. Evaluation datasets do not store model weights or replace production databases. Their primary purpose is systematic and repeatable quality assessment.

Question 91. What is hybrid search designed to combine?

1. Model training and model deployment
   2. Keyword-based and semantic retrieval
   3. Authentication and authorization
   4. Monitoring and billing

Correct Answer: 2. Keyword-based and semantic retrieval

Explanation:

Hybrid search combines lexical or keyword-based retrieval with semantic retrieval to take advantage of the strengths of both approaches. Keyword search can perform well when exact terms, identifiers, product codes, names, or technical phrases are important. Semantic search can identify conceptually related content even when the wording differs between a query and a document. Combining the two can improve retrieval robustness across different types of user questions and source material. Hybrid search is therefore particularly useful for knowledge bases containing both natural-language content and exact terminology. Authentication, monitoring, and billing are separate operational concerns and are not components that hybrid search is intended to combine.

Question 92. Which change can reduce irrelevant information being passed to a generation model?

1. Retrieve every document in the index
   2. Remove all relevance scoring
   3. Use appropriate top-k selection and filtering
   4. Increase the temperature substantially

Correct Answer: 3. Use appropriate top-k selection and filtering

Explanation:

Appropriate top-k selection and filtering can reduce the amount of irrelevant information passed from the retrieval system to the generation model. Instead of supplying a large collection of loosely related documents, the application can select a smaller set of highly relevant results and apply filters based on metadata or other constraints. This helps the model focus on useful evidence and can reduce context-window pressure. Retrieving every document can introduce noise and unnecessary tokens. Removing relevance scoring makes it harder to identify useful results, while temperature affects generation behavior rather than retrieval. Carefully controlling retrieved context is therefore an important part of building an effective RAG pipeline.

Question 93. What is an important reason to monitor latency in a production generative AI application?

1. Latency determines the application’s response time experienced by users
   2. Latency automatically improves factual accuracy
   3. Latency replaces evaluation datasets
   4. Latency permanently changes model weights

Correct Answer: 1. Latency determines the application’s response time experienced by users

Explanation:

Latency measures how long an application or one of its components takes to respond to a request. Monitoring latency is important because slow responses can negatively affect user experience and may indicate bottlenecks in retrieval, model inference, networking, database operations, or application logic. In a RAG application, teams may monitor separate stages to determine whether delays come from vector search, reranking, prompt construction, model generation, or another service. Latency does not directly measure factual accuracy and cannot replace evaluation datasets. Tracking it alongside quality, error, and throughput metrics gives teams a broader view of production performance and helps identify operational problems.

Question 94. What is the main purpose of rate limiting for a model-serving endpoint?

1. Increase hallucination frequency
   2. Remove access controls
   3. Prevent embeddings from being generated
   4. Control request volume and protect service capacity

Correct Answer: 4. Control request volume and protect service capacity

Explanation:

Rate limiting controls how many requests a user, application, or client can send to a service within a defined period. For generative AI endpoints, this can help prevent sudden traffic spikes or excessive usage from overwhelming available compute resources. Rate limits can also support fair resource allocation, cost management, and predictable service behavior. When a limit is exceeded, the application can return an appropriate response or encourage the client to retry later according to defined policies. Rate limiting does not remove access controls or directly improve model reasoning. It is primarily an operational mechanism for controlling demand and protecting service capacity.

Question 95. Which practice improves reproducibility when evaluating different generative AI configurations?

1. Change multiple components without recording them
   2. Record model, prompt, retrieval, and evaluation configuration
   3. Delete previous evaluation results
   4. Use different test questions for every experiment

Correct Answer: 2. Record model, prompt, retrieval, and evaluation configuration

Explanation:

Reproducible evaluation requires recording the important configuration used during an experiment. For a generative AI application, this may include the model version, prompt template, retrieval configuration, embedding model, chunking strategy, evaluation dataset, and relevant generation parameters. Keeping these details allows developers to understand why one experiment produced different results from another and makes successful configurations easier to reproduce. Changing several components without recording them makes it difficult to identify the cause of improvements or regressions. Deleting prior results and changing test questions for every experiment also prevents meaningful comparisons. Proper experiment tracking supports controlled development and reliable evaluation.

Question 96. What is a common benefit of caching in a generative AI application?

1. It can reduce repeated computation and improve response latency
   2. It guarantees that model responses are factually correct
   3. It eliminates the need for access control
   4. It permanently expands the model context window

Correct Answer: 1. It can reduce repeated computation and improve response latency

Explanation:

Caching stores reusable results so that repeated or equivalent requests do not always require the same expensive computation. In generative AI applications, caching may be useful for repeated retrieval queries, embeddings, or other deterministic and reusable operations, depending on the application’s requirements. This can reduce processing time, lower resource consumption, and improve response latency. However, cached information must be managed carefully because stale data can become inaccurate when the underlying knowledge source changes. Caching does not guarantee factual correctness, replace authorization, or expand a model’s context window. Its main operational value is improving efficiency by avoiding unnecessary repeated work.

Question 97. Why can source-document quality strongly affect RAG answer quality?

1. The generation model always ignores retrieved documents
   2. Poor source information can lead to poor or unsupported responses
   3. Source quality only affects network speed
   4. High-quality documents eliminate the need for retrieval

Correct Answer: 2. Poor source information can lead to poor or unsupported responses

Explanation:

RAG applications depend on retrieved information to provide evidence and context for generation. If the source documents are outdated, incomplete, contradictory, duplicated, poorly structured, or factually incorrect, retrieval may return information that does not adequately support the user’s question. Even a strong language model may produce a poor response when the evidence supplied to it is low quality. Maintaining source quality therefore includes processes such as removing obsolete content, correcting inaccurate documents, preserving useful metadata, and ensuring that important information is accessible for retrieval. Good source data does not guarantee perfect answers, but it provides a stronger foundation for grounded generation.

Question 98. What is the role of a system prompt in a generative AI application?

1. It stores vector embeddings
   2. It performs database indexing
   3. It defines high-level instructions and behavioral constraints
   4. It replaces application monitoring

Correct Answer: 3. It defines high-level instructions and behavioral constraints

Explanation:

A system prompt provides high-level instructions that influence how a generative AI model should behave during an interaction. It can establish the assistant’s role, response requirements, constraints, tone, formatting expectations, and other behavioral rules. In a domain-specific application, it may instruct the model to rely on supplied context, avoid unsupported claims, or follow a particular output structure. A system prompt is not responsible for storing vector embeddings or indexing a database, and it does not replace application monitoring. Its purpose is to establish guidance for model behavior and provide consistent instructions that can be combined with user input and retrieved context.

Question 99. What should an application do when retrieved context does not contain enough information to answer a question reliably?

1. Automatically invent missing facts
   2. Ignore the retrieved context
   3. Increase temperature until an answer appears
   4. Acknowledge insufficient information or request clarification

Correct Answer: 4. Acknowledge insufficient information or request clarification

Explanation:

When retrieved context does not contain enough reliable information, the application should avoid presenting unsupported claims as facts. Depending on the design, the assistant can state that the available information is insufficient, ask the user for clarification, request additional information, or direct the user toward an appropriate source. This behavior helps reduce hallucinations and preserves user trust. Simply increasing temperature does not provide missing evidence and may make outputs less predictable. Ignoring retrieved context removes an important grounding mechanism, while inventing facts creates unsupported information. A well-designed generative AI application should recognize evidence limitations rather than attempting to fill gaps with speculation.

Question 100. Which approach best supports continuous improvement of a production RAG application?

1. Deploy once and never measure results
   2. Change the model randomly whenever responses seem poor
   3. Monitor production behavior, evaluate changes, and iterate systematically
   4. Remove user feedback from the development process

Correct Answer: 3. Monitor production behavior, evaluate changes, and iterate systematically

Explanation:

Continuous improvement requires an organized feedback and evaluation cycle rather than uncontrolled changes. A production RAG application can be monitored for quality, latency, errors, retrieval behavior, usage patterns, and other relevant signals. New prompts, retrieval strategies, model versions, or data-processing changes can then be evaluated against representative datasets before being introduced into production. User feedback can provide additional information about real-world failures and unmet needs when collected and handled appropriately. Systematic iteration makes it easier to identify regressions, reproduce successful configurations, and roll back problematic changes. This approach supports reliable improvement while maintaining visibility into application behavior over time.