Implement information extraction solutions 5.1 Build retrieval and grounding pipelines
With integrated vectorization in Azure AI Search, users submit plain-text questions and…
Study guide skill: Configure semantic search, hybrid search, and vector search for grounding
Question B1-046
With integrated vectorization in Azure AI Search, users submit plain-text questions and the service must convert them to vectors at query time. What do you configure, and what constraint applies?
Select one answer.
Answer: C
A vectorizer is defined in the index schema, assigned to a vector field, and used automatically at query time to turn a text query into a vector. It must match the embedding model used to encode the content. Text Split is an indexing-time chunking skill, and semantic configurations or scoring profiles don't vectorize queries.
Deep dive
How it works
Integrated vectorization extends both Azure AI Search pipelines: it encodes vectors during indexer-driven indexing and during queries. The query side is what this question is about. To turn a plain-text question into a vector at query time, you need:
- a query that targets one or more vector fields,
- a text string to convert, and
- a vectorizer, defined in the index schema and assigned to a vector field through a vector profile. Search uses it automatically at query time.
The key constraint is that the vectorizer must match the embedding model used to encode your content. The skill creates embeddings during indexing and the vectorizer creates them at query time, so choose a vectorizer and skill pair that point to the same model deployment. For example, the AzureOpenAIEmbedding skill pairs with the Azure OpenAI vectorizer, and a custom skill pairs with the Custom Web API vectorizer.
On the indexing side you need an indexer, a target index, and a skillset with a chunking strategy (Text Split, the Content Understanding skill or a parsing mode) plus an embedding skill. Vector conversion only goes one way: there's no way to turn vectors back into text, which is why indexes keep both vector and nonvector fields. Integrated vectorization is available in all regions and tiers.
The other options, and when you'd use them
- A scoring profile: Scoring profiles boost or suppress the ranking of matching documents using weighted fields or functions such as freshness, distance and magnitude. Use one to tune relevance.
- A semantic configuration: This sets up the index for semantic reranking of string (nonvector) fields. Use it to add L2 reranking, captions and answers.
- A Text Split skill on the query: Text Split breaks documents into sentence or page chunks during indexing, which is useful when downstream embedding skills have input limits. It's non-billable and runs in the skillset, not on queries.
Exam tip
"Plain-text query against vector fields" means a vectorizer on the field's vector profile, using the same embedding model as the indexing skill. Chunking and embedding documents is the skillset's job.
References
Last reviewed: