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

 

Question 321. Which component determines how relevant documents are selected from a knowledge base for a user’s query?

1. Retriever
   2. Output validator
   3. Response formatter
   4. Application logger

Correct Answer: 1. Retriever

Explanation:

The retriever is responsible for finding information from the application’s knowledge base that is relevant to a user’s query. Depending on the architecture, retrieval can use vector similarity search, keyword search, hybrid search, metadata filters, or a combination of these techniques. The quality of retrieval directly affects the quality of a RAG response because the language model can only effectively ground its answer in information that is supplied to it. A retriever may also work with a reranker to improve the ordering of candidate documents. Retrieval quality should be evaluated independently using metrics such as precision and recall.

Question 322. What is the main purpose of grounding instructions in a RAG prompt?

1. To increase the model’s parameter count
   2. To encourage the model to base its response on supplied evidence
   3. To remove the retrieval stage
   4. To increase embedding dimensions

Correct Answer: 2. To encourage the model to base its response on supplied evidence

Explanation:

Grounding instructions tell the language model how it should use the retrieved information when generating a response. A well-designed prompt can instruct the model to rely on the provided context, avoid unsupported claims, and acknowledge when the available evidence is insufficient. This is particularly important in RAG applications because retrieval provides external information that should guide generation. Grounding instructions do not guarantee that a model will always produce a correct answer, so evaluation remains necessary. They work together with high-quality retrieval, appropriate context assembly, source traceability, and application-level controls to improve the reliability of generated responses.

Question 323. Which factor should be considered when selecting an embedding model for an enterprise RAG application?

1. Only the model’s display name
   2. The application’s language and domain requirements
   3. The user’s screen resolution
   4. The number of application buttons

Correct Answer: 2. The application’s language and domain requirements

Explanation:

Embedding model selection should consider the characteristics of the application’s documents and queries. Language coverage, domain terminology, semantic retrieval quality, embedding dimensions, latency, and operational requirements can all influence the choice. An embedding model that performs well on general-purpose content may not perform equally well on specialized terminology, multilingual content, or technical documentation. The model should therefore be evaluated using representative queries and documents from the intended application. Changing an embedding model can also require re-embedding the indexed content so that document and query representations remain compatible. Evaluation should measure whether the new model improves retrieval quality for the actual use case.

Question 324. Why is it useful to evaluate retrieval quality separately from final answer quality?

1. It helps identify whether failures originate in retrieval or generation
   2. It eliminates the need for model evaluation
   3. It guarantees perfect responses
   4. It prevents documents from being updated

Correct Answer: 1. It helps identify whether failures originate in retrieval or generation

Explanation:

A RAG response depends on multiple stages, so evaluating only the final answer can make troubleshooting difficult. A poor response may occur because the correct document was never retrieved, because irrelevant documents were supplied, or because the language model failed to use correct retrieved evidence. Evaluating retrieval separately with metrics such as precision and recall helps identify problems in the search stage. Final answer evaluation can then focus on qualities such as groundedness, correctness, relevance, and completeness. Separating these evaluations gives development teams clearer evidence about which component needs improvement and reduces the risk of changing the wrong part of the system.

Question 325. What is a potential disadvantage of retrieving too many documents for a single query?

1. It always improves response accuracy
   2. It removes token consumption
   3. It can introduce irrelevant context and increase cost or latency
   4. It prevents semantic search

Correct Answer: 3. It can introduce irrelevant context and increase cost or latency

Explanation:

Retrieving a very large number of documents may increase recall, but it can also introduce irrelevant or redundant information into the model’s context. More context can increase token usage, inference latency, and operational cost. It may also make it harder for the model to identify the most important evidence, especially when several documents contain overlapping or conflicting information. A suitable top-k value should therefore be selected through evaluation. Metadata filtering, reranking, similarity thresholds, and query rewriting can further improve the quality of the final context. The goal is not to retrieve the maximum possible number of documents but to provide sufficient high-quality evidence.

Question 326. Which technique can help a RAG system prioritize newer information when multiple document versions exist?

1. Temperature adjustment
   2. Effective-date or version metadata filtering
   3. Increasing embedding dimensions
   4. Removing source identifiers

Correct Answer: 2. Effective-date or version metadata filtering

Explanation:

When a knowledge base contains multiple versions of the same document, metadata can help identify which version should be considered authoritative. Effective dates, publication dates, status fields, and version identifiers can be used to filter or rank retrieved documents. For example, an application can restrict retrieval to documents that are currently effective rather than returning historical policies. This approach prevents the model from receiving obsolete evidence and reduces ambiguity during generation. Metadata should be maintained as part of the ingestion process and applied consistently during retrieval. Version-aware retrieval is particularly important for policies, procedures, product documentation, and other information that changes over time.

Question 327. What does a reranker typically receive as input?

1. Only the final generated answer
   2. Candidate documents and the user’s query
   3. Only application logs
   4. The entire foundation model

Correct Answer: 2. Candidate documents and the user’s query

Explanation:

A reranker typically receives a user query along with a set of candidate documents that were initially retrieved. It evaluates the relationship between the query and each candidate and produces a more refined relevance ordering. This allows the system to select the strongest evidence before constructing the context sent to the language model. The initial retrieval stage is often optimized for speed and broad candidate discovery, while reranking can perform a more detailed relevance assessment on a smaller set of candidates. Reranking therefore acts as an additional quality layer between initial retrieval and final context construction in many RAG architectures.

Question 328. Which practice helps ensure that only authorized users can retrieve restricted enterprise documents?

1. Authorization-aware retrieval filtering
   2. Higher model temperature
   3. Larger chunk sizes
   4. Removing document metadata

Correct Answer: 1. Authorization-aware retrieval filtering

Explanation:

Authorization-aware retrieval ensures that access permissions are enforced before restricted documents are provided as context to a language model. Documents can contain metadata describing ownership, user groups, roles, classifications, or other access attributes. The retrieval process can use these attributes to limit candidate documents to those the requesting user is allowed to access. This is safer than relying solely on the language model to avoid revealing restricted information after it has already received that information. Access control should be consistently enforced across the application and data layers. Security testing should also verify that users cannot bypass retrieval restrictions through alternative query paths.

Question 329. What is one reason to preserve document identifiers during the ingestion process?

1. They increase model temperature
   2. They support traceability and source citation
   3. They remove the need for embeddings
   4. They guarantee retrieval recall

Correct Answer: 2. They support traceability and source citation

Explanation:

Document identifiers provide a link between retrieved chunks and their originating source documents. This can support citations, auditing, debugging, and investigation of generated responses. When a model produces an answer, the application can use source identifiers to show users which documents contributed evidence to the response. Engineers can also use these identifiers to determine whether a problematic answer resulted from an incorrect source, stale content, poor chunking, or retrieval errors. Identifiers do not make information automatically accurate, but they provide an important traceability mechanism. They should be preserved through ingestion, indexing, retrieval, and response-generation stages where appropriate.

Question 330. Which situation is most likely to require incremental ingestion rather than rebuilding an entire knowledge base?

1. Only a small number of source documents have changed
   2. The application has no documents
   3. The model’s temperature changes
   4. A user changes their screen size

Correct Answer: 1. Only a small number of source documents have changed

Explanation:

Incremental ingestion allows a system to process only newly added, modified, or removed documents instead of rebuilding the entire knowledge base. This can reduce processing time, compute requirements, and operational cost when changes are relatively small compared with the total dataset. Techniques such as document hashes, timestamps, version identifiers, and source-system change tracking can help identify which records require processing. Modified documents may need new chunks and embeddings, while removed documents should be removed from the retrieval index. Incremental ingestion is particularly useful for large enterprise knowledge bases that receive frequent updates and require timely synchronization with authoritative source systems.

Question 331. What is a common purpose of a similarity threshold in a RAG retrieval pipeline?

1. To decide whether retrieved evidence is sufficiently relevant
   2. To determine the model’s number of layers
   3. To create user accounts
   4. To replace access control

Correct Answer: 1. To decide whether retrieved evidence is sufficiently relevant

Explanation:

A similarity threshold establishes a minimum level of similarity that retrieved content must achieve before it is treated as relevant evidence. This can help prevent weakly related documents from being included in the model’s context. If no document meets the threshold, the application can avoid generating a strongly grounded answer and instead request clarification or explain that sufficient evidence was not found. Threshold selection should be based on evaluation data because different embedding models and domains can produce different similarity distributions. A threshold that is too high may exclude useful information, while a threshold that is too low may allow irrelevant content into the context.

Question 332. Why is source-document quality important for RAG performance?

1. Poor source data can lead to poor retrieval and unsupported responses
   2. Source quality only affects the user interface
   3. The language model automatically corrects all source errors
   4. Source quality has no relationship to retrieval

Correct Answer: 1. Poor source data can lead to poor retrieval and unsupported responses

Explanation:

A RAG system depends on the quality of the information available in its knowledge base. If source documents contain outdated, duplicated, incomplete, contradictory, or incorrect information, retrieval may successfully find those documents but still provide poor evidence to the language model. Data preparation should therefore include processes for identifying obsolete material, resolving duplicates, preserving authoritative versions, and maintaining useful metadata. Source quality should be considered before focusing exclusively on model or prompt improvements. Even an effective retrieval architecture cannot reliably produce grounded answers when the underlying knowledge base does not contain accurate and appropriate information.

Question 333. Which method can help identify whether a new RAG configuration actually improves performance?

1. Compare it with a documented baseline using the same evaluation dataset
   2. Deploy it without testing
   3. Change the evaluation dataset for each version
   4. Ignore retrieval metrics

Correct Answer: 1. Compare it with a documented baseline using the same evaluation dataset

Explanation:

A documented baseline provides a stable reference for evaluating changes to a RAG system. When a new configuration is tested against the same representative dataset and metrics, developers can determine whether retrieval, groundedness, relevance, latency, or other measurements have improved or regressed. Keeping the evaluation conditions consistent is important because changing both the configuration and test dataset makes results difficult to interpret. Configuration versions should also be recorded so that successful experiments can be reproduced. Baseline comparisons are particularly useful when testing new prompts, embedding models, retrieval settings, rerankers, or model versions.

Question 334. What should a RAG system do when retrieved evidence does not support a requested claim?

1. Generate a confident answer anyway
   2. Increase temperature
   3. State that the available evidence is insufficient
   4. Invent a citation

Correct Answer: 3. State that the available evidence is insufficient

Explanation:

A grounded RAG system should distinguish between information supported by retrieved evidence and information that is not supported. If the retrieved documents do not contain sufficient evidence for a requested claim, the application should avoid presenting an unsupported statement as factual. Depending on the use case, it can explain that the available sources do not provide enough information, ask the user to clarify the request, or direct the user to an appropriate source. This behavior reduces hallucination risk and makes the application’s limitations transparent. Groundedness evaluation and appropriate prompt instructions can help verify that the system follows this behavior consistently.

Question 335. What is the main purpose of monitoring latency in a production generative AI application?

1. To determine whether responses meet performance expectations
   2. To increase document quality
   3. To change the embedding model automatically
   4. To eliminate authorization checks

Correct Answer: 1. To determine whether responses meet performance expectations

Explanation:

Latency monitoring measures how long the application takes to process requests and return responses. In a RAG system, latency can come from several stages, including query processing, embedding generation, retrieval, reranking, prompt construction, model inference, and downstream validation. Monitoring these stages can help identify bottlenecks and determine whether performance meets application requirements. A sudden increase in latency may indicate changes in retrieval volume, model behavior, infrastructure, or other components. Latency should be considered alongside quality and cost because reducing response time by removing necessary retrieval or validation steps could negatively affect reliability.

Question 336. Which approach can reduce repeated embedding computation for documents that have not changed?

1. Re-embed every document on every request
   2. Cache embeddings and detect unchanged documents
   3. Disable document identifiers
   4. Increase model temperature

Correct Answer: 2. Cache embeddings and detect unchanged documents

Explanation:

Caching embeddings can prevent unnecessary recomputation for documents whose content has not changed. A system can use identifiers, content hashes, timestamps, or version information to determine whether a source document has changed since its previous ingestion. Unchanged documents can retain their existing embeddings, while modified documents can be reprocessed. This approach can reduce ingestion time, computational cost, and resource consumption, particularly for large knowledge bases. Cache invalidation remains important because an embedding should be regenerated when the underlying content or embedding model changes. The ingestion pipeline should therefore track enough metadata to determine when cached representations remain valid.

Question 337. Why can hybrid retrieval be valuable for technical documentation?

1. It combines semantic understanding with exact term matching
   2. It eliminates the need for document indexing
   3. It guarantees correct model outputs
   4. It removes all metadata requirements

Correct Answer: 1. It combines semantic understanding with exact term matching

Explanation:

Technical documentation often contains both natural-language concepts and exact identifiers such as error codes, command names, product versions, configuration parameters, and API names. Semantic search can identify conceptually related content even when wording differs, while lexical search can provide strong results for exact terms. Hybrid retrieval combines these strengths and can improve coverage across different query types. For example, a user asking about a specific error code may benefit from keyword matching, while a conceptual question may benefit more from semantic similarity. Evaluation should determine how retrieval methods and ranking strategies perform on the application’s representative technical queries.

Question 338. What is an important reason to use the same evaluation dataset when comparing prompt versions?

1. It ensures the comparison uses consistent test conditions
   2. It prevents prompt changes
   3. It eliminates the need for metrics
   4. It makes every prompt produce identical output

Correct Answer: 1. It ensures the comparison uses consistent test conditions

Explanation:

Using the same evaluation dataset provides a controlled basis for comparing different prompt versions. If each prompt is evaluated against different questions, changes in results may be caused by differences in the test data rather than by the prompt itself. A stable evaluation dataset allows developers to measure changes in response quality, groundedness, relevance, formatting, or other defined metrics under comparable conditions. The dataset should be representative of the application’s expected workload and should be protected from accidental contamination when possible. Prompt versions should also be recorded so that successful configurations can be reproduced or restored later.

Question 339. Which action can improve RAG retrieval when a user’s question is too broad?

1. Retrieve every document
   2. Ask for clarification or rewrite the query into a more specific form
   3. Increase temperature
   4. Remove all filters

Correct Answer: 2. Ask for clarification or rewrite the query into a more specific form

Explanation:

Broad queries can produce retrieval results covering multiple topics, which may make it difficult for the model to identify the evidence most relevant to the user’s actual intent. Query rewriting can make the request more specific by incorporating known conversation context, important entities, or missing terminology. If the system cannot determine the intended scope reliably, asking a clarification question may be more appropriate. Retrieving every document is inefficient and can overwhelm the context with irrelevant information. Improving query specificity helps the retrieval stage focus on the information that is most likely to answer the user’s actual question.

Question 340. Which practice best supports safe deployment of a new RAG configuration?

1. Deploy directly without evaluation
   2. Remove the previous configuration
   3. Evaluate the new version, monitor deployment behavior, and retain a rollback option
   4. Disable logging during deployment

Correct Answer: 3. Evaluate the new version, monitor deployment behavior, and retain a rollback option

Explanation:

Safe deployment requires more than confirming that the application starts successfully. A new RAG configuration should first be evaluated against representative test data to identify changes in retrieval quality, groundedness, relevance, latency, cost, and other important metrics. After deployment, production monitoring can identify unexpected behavior that was not visible during testing. Retaining a previously validated configuration provides a practical rollback path if the new version causes unacceptable problems. Versioning prompts, models, embedding configurations, retrieval settings, and related artifacts makes rollback more reliable. This controlled process helps teams improve the application while maintaining operational stability and traceability.