AI-102 Azure AI Engineer Associate - Set 3 - Part 1

Correct Answer: XML, JSON, YAML

Retrieval Augmented Generation allows the model to fetch and use external, current data dynamically, making it ideal for frequently updated information without requiring retraining.

Option a is incorrect because fine-tuning weekly is resource-intensive, time-consuming, and impractical for daily updates, making it an inefficient solution for rapidly changing content.

Option c is incorrect because increasing context window size does not solve the problem of accessing external or updated data that exists outside the model's training set.

Option d is incorrect because deploying multiple model versions does not address the need to access real-time or frequently updated external information for generating accurate responses.

Correct Answer: Inserting text into context window

RAG retrieves relevant external data and directly inserts it into the model's context window as raw text, allowing the model to use it for response generation.

Option a is incorrect because RAG does not require retraining or embedding data during the training phase; it operates by retrieving information at inference time instead.

Option b is incorrect because RAG does not store private data in model weights; it retrieves external information dynamically when needed, keeping the model separate from the data.

Option d is incorrect because RAG does not modify the model's architecture or create custom layers; it simply augments the input with retrieved external information.

Correct Answer: Use the identical embedding model

The same embedding model must be used for both storing and searching data to ensure vector compatibility and accurate similarity matching in the retrieval process.

Option a is incorrect because using different embedding models for storage and search would create incompatible vector spaces, resulting in poor or meaningless search results and retrieval accuracy.

Option c is incorrect because version consistency matters more than using the latest version; mismatched versions of embedding models can produce incompatible vectors for effective searching.

Option d is incorrect because using multiple embedding models simultaneously would create inconsistent vector representations, making it impossible to accurately match queries with stored data.

Correct Answer: Embedding Model

The Embedding Model is specifically responsible for transforming text data into vectors, which are numerical representations that enable semantic similarity matching in the system.

Option a is incorrect because the Vector Store is a database that holds the already-created vectors; it does not perform the conversion of text to numerical representations.

Option c is incorrect because the Search Engine compares and retrieves vectors based on similarity; it does not create the numerical representations from the original text data.

Option d is incorrect because the Language Model generates responses based on retrieved information; it does not handle the conversion of text into the numerical vector representations.

Correct Answer: Vector database like MongoDB

Vector databases are specifically optimized for storing vector data and performing efficient similarity searches, making them ideal for RAG systems that require fast vector retrieval.

Option a is incorrect because relational databases are designed for structured tabular data with SQL queries, not for the high-dimensional vector similarity searches required in RAG systems.

Option b is incorrect because document databases are optimized for storing and retrieving JSON-like documents, not for the specialized vector similarity computations needed in RAG architectures.

Option d is incorrect because graph databases are designed for storing and querying relationship-based data structures, not for performing efficient vector similarity searches on embedded representations.

Correct Answer: Vectorize the user query

The user query must first be converted into a vector using the embedding model so it can be compared against the vectors stored in the database.

Option a is incorrect because the system cannot generate a response without first retrieving relevant information from the vector store that matches the user's query intent.

Option c is incorrect because retrieving all documents would be inefficient and unnecessary; the system needs to vectorize the query first to identify which documents are actually relevant.

Option d is incorrect because ranking results can only happen after relevant documents are retrieved, and retrieval requires the query to be vectorized first for comparison purposes.

Correct Answer: Models have specific modalities

Document cracking is necessary because language models have specific modalities or senses; monomodal models can only understand text, requiring non-text formats to be converted appropriately.

Option b is incorrect because while vectors do require preprocessing, this does not explain the fundamental reason why different file types like PDFs and audio need to be converted.

Option c is incorrect because database input requirements are not the primary reason for document cracking; the issue is the model's ability to understand different data formats.

Option d is incorrect because embedding normalization is a separate technical process and does not address the core issue of converting various file formats into model-readable text.

Correct Answer: Multimodal model

Multimodal models can process multiple data types including images, which allows them to directly interpret scanned documents as images without requiring text extraction through parsing.

Option a is incorrect because monomodal text models can only understand plain text input, so they would require a parser to extract text from scanned PDF images first.

Option c is incorrect because embedding models convert data into vectors but do not have the capability to directly interpret image-based documents without prior text extraction.

Option d is incorrect because retrieval models are designed to search and match information, not to interpret or read the content of scanned documents in image format.

Correct Answer: Hierarchical Navigable Small World

HNSW is specifically designed for efficient approximate nearest neighbor searches in vector databases, using graph-based hopping to quickly find similar vectors without exhaustive scanning.

Option a is incorrect because Binary Search Trees are designed for ordered data lookups, not for finding similar vectors in high-dimensional space used in vector databases.

Option b is incorrect because Hash Table Lookups work for exact matches but cannot efficiently find similar or approximate matches in high-dimensional vector spaces required for semantic search.

Option d is incorrect because Linear Sequential Scan would examine every record in the database, which is extremely inefficient and impractical for large-scale vector search operations.

Correct Answer: Semantic ranking

Semantic ranking analyzes the contextual meaning and intent behind queries to re-order results, ensuring that documents matching the user's actual intent are prioritized over simple keyword matches.

Option a is incorrect because keyword filtering would still rely on matching specific words rather than understanding the contextual meaning or intent behind the user's query about capital.

Option c is incorrect because alphabetical sorting organizes results by letter order, which has no relationship to the semantic meaning or relevance of documents to the user's intent.

Option d is incorrect because frequency analysis counts word occurrences but does not distinguish between different meanings of the same word or understand the query's contextual intent.

Correct Answer: Hybrid search

Hybrid search combines vector search with traditional text search, allowing the system to match both semantic similarity and exact keywords, improving overall coverage and retrieval effectiveness.

Option a is incorrect because pure vector search relies only on semantic embeddings and may miss documents with exact keyword matches that use different but relevant terminology.

Option b is incorrect because full text search only uses keyword matching and would miss semantically similar documents that use entirely different words with the same meaning.

Option d is incorrect because filtered search applies constraints to narrow results but does not combine different search methodologies to improve coverage and capture diverse relevant documents.

Correct Answer: Semantic hybrid search

Semantic hybrid search combines vector and text search with an additional semantic re-ranking step, typically producing the best results by leveraging multiple retrieval and ranking strategies.

Option a is incorrect because vector search alone only uses embeddings for similarity matching and misses the benefits of keyword matching and semantic re-ranking for optimal results.

Option b is incorrect because full text search relies solely on keyword matching and lacks the semantic understanding and vector similarity capabilities needed for the highest quality retrieval.

Option c is incorrect because while hybrid search combines vector and keyword matching, it lacks the semantic re-ranking layer that provides the contextual understanding for truly optimal results.

Correct Answer: Indexer

The Indexer is the automated ETL process that crawls data sources, extracts information, and populates the Index, and it can run on scheduled intervals for continuous updates.

Option a is incorrect because the Data Source is simply the origin location of the data, such as Azure Blob Storage, not the component that actively crawls and extracts information.

Option c is incorrect because Skillsets are AI enrichments that transform data during indexing, but they do not perform the actual crawling and extraction of data from sources.

Option d is incorrect because the Search Engine executes queries and retrieves results from the index; it does not handle the data extraction and population process from external sources.

Correct Answer: Index

The Index is the physical store within Azure AI Search that holds references to documents along with their vectorized and chunked data for retrieval operations.

Option a is incorrect because the Data Source is the origin of raw documents like Azure Blob Storage, not the location where processed, vectorized data is stored.

Option b is incorrect because the Indexer is the process that moves and transforms data, not the storage location where the vectorized and chunked information actually resides.

Option d is incorrect because Skillsets are transformation rules applied during indexing to enrich data, not the physical storage component that holds the processed vector data.

Correct Answer: Enrich with AI skills for OCR

Enabling AI Skills with OCR during setup allows Azure AI Services to perform Optical Character Recognition on images and scanned documents, extracting text for indexing.

Option a is incorrect because vector search profiles configure the similarity search algorithm but do not handle the extraction of text from images or scanned documents.

Option b is incorrect because semantic ranking re-orders search results based on contextual understanding but does not extract text from images or process scanned PDF files.

Option d is incorrect because hybrid search configuration combines different search methods but does not provide the image text extraction capability needed for scanned documents.

Correct Answer: Generate embeddings from text

Raw text must be converted into vector embeddings using an embedding model before it can be uploaded and searched through vector similarity matching in the system.

Option a is incorrect because uploading raw documents directly without vectorization would not enable vector similarity searches; the text must first be transformed into numerical vector representations.

Option c is incorrect because configuring search filters helps narrow results based on metadata but does not address the fundamental requirement of converting text into searchable vector representations.

Option d is incorrect because enabling semantic ranking improves result ordering but does not create the vector embeddings necessary for the initial similarity-based search functionality.

Correct Answer: HNSW

HNSW algorithm is specifically designed for vector databases and is configured in the vector search profile to enable efficient approximate nearest neighbor searches in high-dimensional space.

Option a is incorrect because Bubble Sort is a basic sorting algorithm for ordered lists, not designed for vector similarity searches or high-dimensional approximate nearest neighbor matching.

Option c is incorrect because Quick Sort is a general-purpose sorting algorithm that operates on comparable elements, not suitable for finding similar vectors in high-dimensional embedding spaces.

Option d is incorrect because Binary Search works on sorted one-dimensional data for exact matches, not for finding approximate nearest neighbors in multi-dimensional vector spaces.

Correct Answer: Vectorizable text query

Vectorizable text queries allow the service to automatically handle the embedding step, converting the user's text query into vectors without requiring manual embedding by the developer.

Option a is incorrect because pure vector search requires the query to already be in vector form, meaning the developer must manually embed the text query before submission.

Option c is incorrect because full text search uses keyword matching without any embedding or vectorization, so it does not handle the embedding step at all.

Option d is incorrect because manual embedding query explicitly requires the developer to embed the text themselves rather than having the service automatically handle the embedding process.

Correct Answer: Before or during vector search

Applying metadata filters before or during vector search reduces the search space, improving performance and ensuring that only relevant category items are considered for vector matching.

Option a is incorrect because applying filters after generating the response wastes computational resources on irrelevant results and provides no performance benefit to the search or retrieval process.

Option c is incorrect because waiting until after semantic ranking means the system has already processed and ranked irrelevant results, making the process inefficient and resource-intensive.

Option d is incorrect because filters during upload do not affect query-time search performance; filters need to be applied when executing searches to narrow results effectively.

Correct Answer: Index and indexer automatically

The Add Your Data feature in Azure OpenAI Studio automatically creates both the Index and Indexer, streamlining the RAG implementation process without requiring manual configuration.

Option a is incorrect because the system creates more than just the index; it also automatically sets up the indexer to crawl and populate the index from the data source.

Option b is incorrect because while it connects to the data source, the feature does much more by automatically creating the index and indexer components for the RAG system.

Option d is incorrect because while skillsets can be configured, the primary automatic creation involves the index and indexer, not just the AI enrichment skillsets.