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

 

Question 301. Which technique can help a RAG system handle user queries containing domain-specific abbreviations?

1. Remove all metadata
   2. Increase temperature
   3. Use query rewriting with domain terminology
   4. Disable retrieval

Correct Answer: 3. Use query rewriting with domain terminology

Explanation:

Query rewriting can improve retrieval when users use abbreviations, acronyms, product codes, or organization-specific terminology that may not directly match the wording in source documents. A rewriting component can expand or clarify the query while preserving the user’s intended meaning. For example, an internal abbreviation could be converted into its full terminology before semantic or hybrid search is performed. This can improve the likelihood that relevant chunks are retrieved. Domain terminology can also be included in query expansion when appropriate. The rewritten query should be evaluated because excessive expansion can introduce unrelated terms and potentially reduce retrieval precision.

Question 302. What is the primary purpose of chunking documents before creating embeddings for a RAG application?

1. To divide large documents into manageable retrieval units
   2. To increase the model’s parameter count
   3. To remove the need for indexing
   4. To guarantee factual responses

Correct Answer: 1. To divide large documents into manageable retrieval units

Explanation:

Chunking divides large source documents into smaller sections that can be independently embedded and retrieved. This allows the retrieval system to identify specific portions of a document that are relevant to a user’s question rather than returning an entire large document. Appropriate chunk boundaries can preserve enough surrounding context while keeping retrieved information focused. Poor chunking can negatively affect retrieval quality: chunks that are too large may contain excessive unrelated information, while chunks that are too small may lose important context. Chunking should therefore be evaluated alongside embedding models, retrieval configuration, and the structure of the underlying documents.

Question 303. Why can overlapping chunks be useful when preparing documents for RAG?

1. They eliminate the need for embeddings
   2. They guarantee that every answer is correct
   3. They increase the foundation model’s training data
   4. They help preserve context that crosses chunk boundaries

Correct Answer: 4. They help preserve context that crosses chunk boundaries

Explanation:

Chunk overlap allows a portion of text from one chunk to appear again in the following chunk. This can help preserve relationships between sentences or concepts that would otherwise be separated by a hard chunk boundary. For example, an explanation may begin near the end of one chunk and continue into the next. Without overlap, retrieving only one of those chunks could provide incomplete context to the language model. However, excessive overlap can create redundant retrieval results and increase storage or token usage. The appropriate overlap size should therefore be selected based on document structure and validated through retrieval and answer-quality evaluation.

Question 304. Which retrieval approach combines semantic similarity with keyword matching?

1. Pure generation
   2. Hybrid search
   3. Fine-tuning
   4. Output validation

Correct Answer: 2. Hybrid search

Explanation:

Hybrid search combines different retrieval methods, commonly semantic vector search and lexical or keyword-based search. Semantic retrieval is useful for identifying content with similar meaning even when the wording differs, while keyword retrieval can be particularly effective for exact terms such as product identifiers, error codes, names, version numbers, and technical phrases. Combining both approaches can provide broader coverage across different query types. The results may then be merged, ranked, or passed through a reranker. Hybrid retrieval is especially useful in enterprise knowledge bases where both natural-language questions and exact identifiers are common.

Question 305. What is the main purpose of an embedding model in a RAG architecture?

1. To convert text into numerical representations suitable for similarity search
   2. To generate the final natural-language response
   3. To enforce user authorization
   4. To validate JSON output

Correct Answer: 1. To convert text into numerical representations suitable for similarity search

Explanation:

An embedding model converts text, such as documents or user queries, into numerical vectors that represent semantic characteristics of the content. These vectors can be stored in a vector index and compared using similarity calculations. During retrieval, the query embedding is compared with document embeddings to identify content that is semantically related to the user’s request. Embeddings are therefore a core component of semantic retrieval but are not responsible for generating the final answer. Choosing an embedding model that fits the language, terminology, and domain of the application can significantly affect retrieval performance and should be evaluated using representative queries.

Question 306. Which metadata field would be most useful for retrieving only the current version of a policy document?

1. Random document color
   2. Number of paragraphs
   3. Effective date or version identifier
   4. Model temperature

Correct Answer: 3. Effective date or version identifier

Explanation:

Effective-date and version metadata allow a retrieval system to distinguish current information from historical or superseded content. Policy documents often exist in multiple versions, and returning an outdated version can result in an incorrect answer even when retrieval technically succeeds. By storing attributes such as version number, publication date, effective date, or status, the application can filter or rank documents according to business rules. This is particularly important when policies change over time. Metadata should be maintained during ingestion and updated when documents change. Combining version-aware filtering with semantic retrieval helps ensure that the model receives evidence that is both relevant and current.

Question 307. What problem can occur when chunks are excessively large?

1. They may contain too much unrelated information
   2. They always improve retrieval precision
   3. They eliminate token usage
   4. They prevent documents from being indexed

Correct Answer: 1. They may contain too much unrelated information

Explanation:

Excessively large chunks can reduce the usefulness of retrieved context because a single retrieved chunk may contain many topics that are unrelated to the user’s question. This increases the amount of information the model must process and can consume valuable context-window capacity. Large chunks may also make it harder for the retrieval system to identify the precise evidence needed for a query. On the other hand, making chunks extremely small can remove important context. Effective chunking therefore requires a balance between focused retrieval and sufficient surrounding information. Document structure, sentence boundaries, headings, and semantic relationships can all help determine appropriate chunk sizes.

Question 308. Why should duplicate or obsolete documents be removed or managed during ingestion?

1. To increase model parameters
   2. To prevent conflicting or redundant information from affecting retrieval
   3. To disable semantic search
   4. To eliminate the need for evaluation

Correct Answer: 2. To prevent conflicting or redundant information from affecting retrieval

Explanation:

Duplicate and obsolete documents can create several retrieval problems. Multiple copies of the same information can consume retrieval slots and reduce the diversity of useful results. Obsolete documents can also conflict with current policies, procedures, or product information. If both old and current versions are retrieved, the language model may have difficulty determining which evidence should be trusted. Ingestion pipelines can address these problems through deduplication, document status metadata, version management, content hashing, and effective-date filtering. Maintaining a clean knowledge base improves retrieval quality and makes the evidence supplied to the model easier to interpret and audit.

Question 309. What is a key purpose of a vector index in a RAG system?

1. To store user passwords
   2. To generate application interfaces
   3. To support efficient similarity searches over embeddings
   4. To replace model evaluation

Correct Answer: 3. To support efficient similarity searches over embeddings

Explanation:

A vector index organizes numerical embeddings so that similarity searches can be performed efficiently. When a user submits a query, the query is converted into an embedding and compared with indexed document embeddings. The vector index helps identify candidate documents whose representations are close to the query according to the selected similarity method. Without an efficient indexing mechanism, searching a large collection of vectors could become computationally expensive. Vector search is only one component of a complete RAG system, however. Metadata filters, reranking, authorization, chunking, and evaluation may also be required to produce reliable retrieval results.

Question 310. What should an application record to make production RAG responses easier to troubleshoot?

1. Only the user’s final screen resolution
   2. Relevant request, retrieval, model, and configuration information
   3. Only the model’s final answer
   4. Nothing, to avoid all observability

Correct Answer: 2. Relevant request, retrieval, model, and configuration information

Explanation:

Effective observability requires enough information to reconstruct how a response was produced. Depending on privacy and security requirements, useful information may include request identifiers, retrieval queries, selected document identifiers, ranking information, prompt or model versions, latency measurements, and output-validation results. This allows engineers to investigate whether a problem originated in data ingestion, retrieval, prompt construction, model generation, or downstream processing. Logs should be designed carefully so that sensitive information is not unnecessarily retained. The goal is not to record everything indiscriminately, but to capture sufficient structured telemetry to diagnose failures, measure performance, and support controlled improvement.

Question 311. Which metric measures the proportion of retrieved documents that are actually relevant to the query?

1. Retrieval precision
   2. Retrieval recall
   3. Model throughput
   4. Generation latency

Correct Answer: 1. Retrieval precision

Explanation:

Retrieval precision measures the proportion of returned retrieval results that are relevant to the user’s query. A low precision value indicates that the retriever is returning many irrelevant documents, which can introduce noise into the model’s context. This can make it harder for the language model to identify the most useful evidence and may increase token usage and latency. Precision can sometimes be improved through better embeddings, metadata filtering, reranking, query rewriting, or a suitable similarity threshold. Precision should be evaluated alongside recall because optimizing only one metric can create trade-offs. A strong retrieval system aims to return useful evidence while avoiding unnecessary irrelevant content.

Question 312. What is the main purpose of a similarity threshold in semantic retrieval?

1. To increase model temperature
   2. To determine whether retrieved items are sufficiently similar to the query
   3. To modify model weights
   4. To create new training data automatically

Correct Answer: 2. To determine whether retrieved items are sufficiently similar to the query

Explanation:

A similarity threshold establishes a minimum relevance level that retrieved documents should meet before they are considered suitable evidence. This can help prevent weakly related documents from being passed to the language model. If no candidate meets the threshold, the application can acknowledge that sufficient evidence was not found or ask the user to clarify the request. The appropriate threshold depends on the embedding model, domain, document collection, similarity method, and desired balance between recall and precision. Thresholds should therefore be evaluated empirically rather than selected arbitrarily. An overly strict threshold can remove useful evidence, while a low threshold may introduce irrelevant context.

Question 313. Which practice can help protect sensitive information from appearing unnecessarily in application logs?

1. Log every piece of retrieved content without filtering
   2. Disable authentication
   3. Apply privacy-aware logging and minimize sensitive data retention
   4. Increase top-k retrieval

Correct Answer: 3. Apply privacy-aware logging and minimize sensitive data retention

Explanation:

Generative AI applications can process sensitive user requests, retrieved documents, and generated responses, so logging must be designed with privacy and security considerations in mind. Applications should avoid recording sensitive content unless it is genuinely required for an operational purpose. Techniques such as redaction, masking, access-controlled logs, data minimization, retention limits, and carefully selected identifiers can reduce unnecessary exposure. Logging policies should also account for who can access diagnostic information and how long it is retained. Strong observability does not mean storing every piece of application data. Instead, teams should capture useful operational signals while minimizing unnecessary exposure of confidential information.

Question 314. Why can a smaller language model sometimes be appropriate for a production RAG application?

1. Smaller models always produce more accurate answers
   2. Smaller models eliminate retrieval requirements
   3. Smaller models cannot hallucinate
   4. A smaller model may reduce latency and cost when it meets quality requirements

Correct Answer: 4. A smaller model may reduce latency and cost when it meets quality requirements

Explanation:

A smaller language model can sometimes provide sufficient quality for a specific RAG task while requiring fewer computational resources than a larger model. This can reduce inference cost and response latency, which may be important for applications with high request volumes or strict response-time requirements. However, model size should not be selected based on cost alone. The model must be evaluated for the application’s required capabilities, grounding behavior, instruction following, structured output requirements, and language or domain needs. If a smaller model performs adequately on representative evaluation data, its operational efficiency may make it a suitable choice for the application.

Question 315. What is the purpose of human evaluation in a generative AI application?

1. To provide qualitative assessment that automated metrics may not fully capture
   2. To replace every automated evaluation
   3. To increase embedding dimensions
   4. To change retrieval results automatically

Correct Answer: 1. To provide qualitative assessment that automated metrics may not fully capture

Explanation:

Human evaluation can identify qualities that are difficult to measure reliably using automated metrics alone. Reviewers may assess whether an answer is useful, clear, appropriately grounded, complete, and aligned with the intended task. Human reviewers can also identify subtle problems such as misleading wording, missing qualifications, inappropriate assumptions, or confusing explanations. Automated metrics remain valuable because they enable repeatable and scalable testing, but human review provides an additional perspective on real-world usefulness. A balanced evaluation process can combine automated measurements with carefully designed human assessments, particularly for important changes, difficult cases, or scenarios where qualitative judgment matters.

Question 316. What should a system do if a user’s query is ambiguous and could refer to multiple topics?

1. Select an arbitrary topic
   2. Increase the temperature
   3. Ask for clarification or use available conversation context to disambiguate
   4. Retrieve the entire knowledge base

Correct Answer: 3. Ask for clarification or use available conversation context to disambiguate

Explanation:

Ambiguous queries can lead to poor retrieval because the system may retrieve documents related to the wrong interpretation of the user’s request. When conversation history provides sufficient information, the application can use that context to rewrite the query more precisely. If the ambiguity cannot be resolved reliably, asking the user for clarification is safer than making an unsupported assumption. Retrieving large amounts of unrelated information does not solve the underlying ambiguity and can increase cost and context usage. Query rewriting, conversation-aware retrieval, and clarification prompts are useful techniques for ensuring that the retrieval stage reflects the user’s actual intent.

Question 317. Which change is most likely to improve retrieval when relevant documents use terminology different from the user’s query?

1. Increase generation temperature
   2. Improve query rewriting or use hybrid retrieval
   3. Remove all metadata
   4. Disable embeddings

Correct Answer: 2. Improve query rewriting or use hybrid retrieval

Explanation:

Terminology mismatch can cause a retrieval system to miss useful documents even when the underlying information is available. Query rewriting can transform a user’s wording into terms that better represent the concepts used in the knowledge base. Hybrid retrieval can also help because keyword search may identify exact terminology while semantic search captures related meaning. These approaches are especially useful in technical and enterprise environments where product names, abbreviations, internal terminology, and formal documentation language may differ from everyday user phrasing. The effectiveness of the change should be measured using representative evaluation queries so that improvements in recall do not introduce excessive irrelevant results.

Question 318. What is an important consideration when updating documents in a production RAG knowledge base?

1. Changes should be reflected through appropriate ingestion and re-indexing processes
   2. Documents should never be updated
   3. Only the language model temperature should change
   4. Existing embeddings should always be reused regardless of content changes

Correct Answer: 1. Changes should be reflected through appropriate ingestion and re-indexing processes

Explanation:

When source documents change, the retrieval index must accurately represent the new content. Depending on the architecture, this can involve detecting modified documents, reprocessing their content, generating new embeddings, updating metadata, and replacing or removing outdated index entries. Simply changing the source file without updating the retrieval system can cause the application to continue returning stale information. Content hashes, version identifiers, timestamps, and incremental ingestion pipelines can help identify which documents need processing. The exact workflow depends on the retrieval architecture, but maintaining synchronization between authoritative source data and the retrieval index is essential for reliable production RAG behavior.

Question 319. Why should prompt, model, embedding, and retrieval configurations be versioned together when evaluating a RAG application?

1. To prevent any future experimentation
   2. To make every output identical
   3. To provide a reproducible record of the system configuration used for a result
   4. To eliminate the need for monitoring

Correct Answer: 3. To provide a reproducible record of the system configuration used for a result

Explanation:

A RAG response can depend on many components beyond the language model itself. Prompt instructions, model versions, embedding models, chunking strategies, retrieval parameters, metadata filters, rerankers, and evaluation configurations can all influence the final result. Versioning these components provides a reproducible record of which configuration produced a particular evaluation or production behavior. This makes experiments easier to compare and failures easier to investigate. It also supports rollback when a newer configuration performs poorly. Without version tracking, teams may know that performance changed but lack enough information to identify which component caused the change.

Question 320. Which approach best supports continuous improvement of a production RAG application?

1. Change multiple components without evaluation
   2. Deploy every experiment immediately
   3. Rely only on occasional user complaints
   4. Monitor production behavior, evaluate controlled changes, and retain validated configurations

Correct Answer: 4. Monitor production behavior, evaluate controlled changes, and retain validated configurations

Explanation:

Continuous improvement should be based on a controlled lifecycle rather than untracked changes. Production monitoring can reveal latency issues, retrieval failures, user feedback patterns, and changes in application behavior. Candidate improvements can then be tested against a stable evaluation dataset using defined metrics. Prompt, model, embedding, retrieval, and data changes should be versioned so that results are reproducible. Successful configurations can be retained as validated versions, while unsuccessful changes can be rolled back or investigated further. This process creates a feedback loop in which real-world observations inform controlled experiments, and evaluation evidence guides which changes are appropriate for future production deployments.