Develop AI Solutions by Using Azure Data Management Services practice questions

Answer: A) The queue's max delivery count, which moves the message to the dead-letter subqueue once exceeded

Service Bus counts delivery attempts on peek-locked messages. When a message goes over the max delivery count, it moves automatically to the dead-letter queue (reason MaxDeliveryCountExceeded), where it stays until someone inspects and settles it.

• ACorrect. Each time a peek-locked message is abandoned or its lock expires, its delivery count goes up. When the count passes the entity's max delivery count (default 10), Service Bus moves the message to the dead-letter queue with reason MaxDeliveryCountExceeded. It stays there until someone receives it and completes it.
• BIncorrect. Duplicate detection drops repeated sends that have the same MessageId within a time window. It does nothing about a message that keeps failing processing, and it doesn't keep the message for review.
• CIncorrect. A longer lock only gives each attempt more time. The document is malformed, so every attempt still fails and the message keeps being redelivered.
• DIncorrect. If dead-lettering on expiration is off, an expired message is simply dropped. The problem document would be lost instead of kept for someone to review.

Answer: A) Set the Category field as facetable

Facets are calculated from each query's result set, and only fields marked facetable can return value buckets with counts. Filterable is usually enabled too, so selecting a bucket can apply a filter.

• ACorrect. A field marked facetable can be named in a query's facets parameter. The service then returns buckets of distinct values with document counts, calculated from the current result set.
• BIncorrect. Filterable lets the app apply a $filter when a user selects a bucket, and the docs recommend enabling it next to facetable. It doesn't make the service compute or return per-value counts.
• CIncorrect. Searchable includes the field in full-text search through lexical analysis. It doesn't produce aggregated value counts.
• DIncorrect. Sortable lets results be ordered by the field's value. It doesn't group values or return counts.

Answer: A) An Azure Monitor metric alert on the Storage usage metric, with a threshold set below the tier's storage capacity limit

Azure AI Search sends a Storage usage (IndexStorageUsage) platform metric to Azure Monitor. A metric alert with a threshold below the tier's limit gives early warning before indexing fails. Failed-document counts, throttling metrics and operation logs don't track storage use against the limit.

• ACorrect. Azure AI Search sends a Storage usage platform metric (IndexStorageUsage) to Azure Monitor. It reports text and vector storage in bytes and can be split by index. An alert with a threshold below the tier's limit warns the team early, so they can add partitions or move to a higher tier before indexing fails.
• BIncorrect. This alert fires only after indexer documents have already failed. That is the reactive discovery the team wants to avoid, and the metric doesn't measure how close the index is to its storage limit.
• CIncorrect. This metric tracks queries that are dropped when request volume exceeds the service's processing capacity. It reflects query load, not storage use against the tier's storage limit.
• DIncorrect. OperationLogs record individual query and indexing operations, such as their duration and result. Collecting them doesn't set up any alert, and the logs don't measure storage use against the tier's limit.

Answer: A) Re-embed the entire document corpus with the new model, re-index it, and vectorize queries with the same new model

Vectors from different embedding models aren't comparable, and Azure AI Search expects the same model to be used for indexing and querying. Switching models means re-embedding and re-indexing the whole corpus, which is a significant, costly operation.

• ACorrect. Vector search compares a query against vectors in one embedding space, and that space is defined by the model that produced the vectors. Microsoft recommends using the same embedding model for indexing and querying. Every existing document vector therefore has to be regenerated with the new model and re-indexed. For a large corpus this can be expensive and should be planned for.
• BIncorrect. Query vectors from the new model would be compared against document vectors from the old model. Different models produce vectors in different spaces, so the similarity scores would be meaningless, and the search would fail outright if the dimensions differ.
• CIncorrect. The similarity metric (cosine, dotProduct, or euclidean) only controls how distances are calculated. Changing it doesn't touch the stored document vectors, which still come from the old model.
• DIncorrect. Rebuilding the HNSW graph with different parameters only reorganizes the vectors already stored. They were all produced by the old model, so the index still reflects the old model's embedding space.

Answer: A) Create and populate products-v2 with the new schema, then update the products alias so its indexes array contains products-v2

For a breaking schema change, build a new index under a new name, then update the alias the application queries so it maps to the new index. This avoids downtime and code changes. An alias maps to one index and can't be an indexer's target.

• ACorrect. An alias is a secondary name that applications use in place of an index name. The documentation says the most common alias update is changing the index name when the underlying index is replaced with a newer version. products-v1 keeps serving queries until the alias is switched. Index aliases are available in the stable REST API (2026-04-01) and the latest stable SDKs.
• BIncorrect. An alias can specify only one index name in its indexes array, so it can't send queries to two indexes.
• CIncorrect. An alias can't be used as the targetIndexName of an indexer. The indexer must target an actual index, such as products-v2.
• DIncorrect. This leaves the application with no usable index while the rebuild runs. Also, you can't delete an index that an alias maps to: the request fails with 400 until the alias is deleted or remapped.

Answer: A) Hash n-gram feature vectors with MinHash (locality-sensitive hashing), then run a similarity join to flag documents above a Jaccard similarity threshold

Exact hashes only match identical content. Microsoft's RAG pipeline guidance recommends MinHash locality-sensitive hashing with a similarity join to catch near-duplicate documents based on their content.

• ACorrect. Microsoft's RAG data pipeline guidance recommends locality-sensitive hashing, specifically MinHash, for content-based deduplication. You build a feature vector for each document from n-grams, hash it with MinHash for Jaccard distance, and run a similarity join to find duplicates and near-duplicates. Then you filter out the copies you don't want to keep.
• BIncorrect. Normalization only removes differences in case, spacing and punctuation. This is still an exact-match check, so if any word differs, the cryptographic hash is completely different and the near-duplicate is missed.
• CIncorrect. The guidance notes that metadata can remove some duplicates but may not be enough. Edited copies can have different titles or dates, and metadata doesn't compare document content, so near-duplicates with small wording changes get through.
• DIncorrect. Semantic chunking decides where chunk boundaries fall based on topic similarity. It doesn't detect or remove near-duplicate documents across the corpus, so redundant chunks from edited copies stay in the index.

Answer: A) A query vector is only meaningfully comparable to document vectors produced by the same model, so mixed vectors give unreliable similarity rankings

Embeddings from different models live in different vector spaces, so a query vector should only be compared with document vectors from the same model. A phased migration needs retrieval logic, such as separate vector fields per model, that accounts for the mixed state.

• ACorrect. Different embedding models produce vectors in different vector spaces. Microsoft advises using the same embedding model for queries as for the indexed documents, and making sure every queried vector field holds embeddings from the same model. During the transition, the team must route each query to vectors from the matching model, for example by using separate vector fields per model.
• BIncorrect. Nearest neighbor search always returns k results, as long as at least k documents exist, even when matches are poor. The real problem is that the rankings are unreliable, not that fewer results come back.
• CIncorrect. Changing the similarity metric doesn't make vectors from different models comparable. Cosine is still the recommended metric for Azure OpenAI embedding models.
• DIncorrect. The keyword part of a hybrid query uses BM25 over text fields, which embeddings don't affect. Only the vector part is affected by mixed models.

Answer: A) Set queryType to full and add the ~ operator after each term

Fuzzy search uses full Lucene syntax (queryType=full) with ~ after each term. It matches indexed terms within a small edit distance (2 by default), so typos still find relevant documents. A ~ after a quoted phrase is proximity search, a different feature.

• ACorrect. In full Lucene syntax (queryType=full), a ~ after a single term starts a fuzzy search. It expands the term to similar terms within a Damerau-Levenshtein edit distance of up to 2 (the default), so "exampel~" can match "example".
• BIncorrect. A ~ after a quoted phrase starts a proximity search, which finds terms within N words of each other. It does not tolerate misspelled characters inside a term.
• CIncorrect. A prefix wildcard such as "exampel*" matches terms that start with the exact characters "exampel". A transposition typo will not match "example".
• DIncorrect. A synonym map expands only the terms that someone explicitly lists in it. Handling typos this way would mean maintaining the list of misspellings that the team wants to avoid.

Answer: A) A container vector policy for the embedding path, a diskANN vector index on that path, and queries that sort results with VectorDistance

Cosmos DB for NoSQL vector search needs a container vector policy and a vector index, and queries use VectorDistance. For high-dimension embeddings searched at large scale, diskANN is the right index; flat is limited to 505 dimensions.

• ACorrect. Vector search requires a container vector policy that defines the path, data type, dimensions and distance function. diskANN supports up to 4,096 dimensions and is generally the best-performing index when a search covers more than 50,000 vectors. Queries use the VectorDistance system function.
• BIncorrect. A flat index supports at most 505 dimensions, so it can't index 1,536-dimension embeddings. It's also a brute-force search, which suits smaller searches better.
• CIncorrect. Spatial indexes and ST_DISTANCE work with geospatial GeoJSON data, not high-dimensional embedding vectors.
• DIncorrect. Composite indexes support ordering and filtering on several scalar properties. They don't calculate similarity between vectors.

Answer: A) It applies the profile named in the index's defaultScoringProfile property; if that property isn't set, no profile is applied

When a query omits scoringProfile, Azure AI Search uses the profile named in the index's defaultScoringProfile property. If that property isn't set, default relevance scoring is used and no profile is applied.

• ACorrect. In the index definition, defaultScoringProfile is the name of the scoring profile to use when a query doesn't specify one. If the property isn't set and the query names no profile, the service uses default relevance scoring without any profile.
• BIncorrect. The order of profiles in the scoringProfiles array doesn't matter. A default exists only when the index's defaultScoringProfile property explicitly names one.
• CIncorrect. A query can use only one scoring profile at a time. functionAggregation controls how multiple scoring functions are combined within a single profile, not across profiles.
• DIncorrect. When a profile was added or modified has no effect on which profile applies. Without a scoringProfile query parameter, only the defaultScoringProfile property decides which profile, if any, is used.