Plan and manage an Azure AI solution 1.1 Choose the appropriate Foundry services for generative AI and agents

You design an Azure AI Search index for long PDF manuals by using integrated…

Study guide skill: Choose an appropriate method for retrieval and indexing

Question B6-003

You design an Azure AI Search index for long PDF manuals by using integrated vectorization. The search UI must show readable passage text for each result. Which two design decisions are correct? (Choose two.)

Select 2 answers.

  • A Assign a vectorizer that uses a cheaper embedding model than the one used at indexing time
  • B Chunk the documents in the skillset (for example, with the Text Split skill) before the embedding skill runs ✓ Correct answer
  • C Keep nonvector text fields next to the vector fields, because vectors can't be converted back into text ✓ Correct answer
  • D Store only vector fields and convert the matched vectors back into text at query time
  • E Skip chunking, because integrated vectorization removes the embedding model's input limits

Answer: B, C

Unless raw documents are small, chunking is needed to meet the embedding model's token input requirements, and the Text Split skill is one of the supported chunking strategies. Vector conversion is one-way, so indexes keep both vector and nonvector fields for readable results. The query-time vectorizer must match the embedding model that encoded the content.

Deep dive

How it works

Integrated vectorization in Azure AI Search adds two capabilities: vector encoding during indexer-driven indexing and vector encoding at query time. The common setup for documents like long PDFs is:

  1. A data source and indexer.
  2. A skillset with a chunking strategy (the Text Split skill, the Azure Content Understanding skill, or a document parsing mode) followed by an embedding skill that vectorizes the chunks.
  3. An index that receives the chunked, vectorized content, with a vectorizer assigned to the vector fields for query time.

Two design rules come straight from the page:

  • Chunk before you embed. Chunking isn't a hard requirement, but unless your raw documents are small, it's necessary to meet the embedding model's token input requirements. Text Split chunking is free.
  • Keep text next to vectors. Vector conversions are one-way (nonvector to vector). There's no vector-to-text conversion for queries or results, which is why indexes contain both vector and nonvector fields. The readable passage text has to be stored as a nonvector field.

The other options, and when you'd use them

  • A cheaper embedding model for the vectorizer: The vectorizer should use the same embedding model that generated the vectors in the index.
  • Store only vectors and convert them back: There's no vector-to-text conversion.
  • Skip chunking: Integrated vectorization doesn't remove the embedding model's input limits. Chunks are how you meet them.

Exam tip

Integrated vectorization = chunk (Text Split) then embed, same model at query time, and always keep the text field.

References

Last reviewed:

Ready to test yourself? The quiz asks this and every other AI-103 question, with instant feedback. Practice this in a quiz →

All Plan and manage questions · AI-103 study guide

Thanks for your feedback!