AI-103 Practice Question 76: Embedding Dimension Tradeoff
Microsoft Azure AI Apps and Agents Developer Associate (AI-103) · Test 8 · question 6 of 10
A search team is indexing 50 million document chunks in Azure AI Search and wants to shrink vector storage as much as possible using binary quantization together with Matryoshka Representation Learning (MRL) dimension truncation. Microsoft's own guidance for this combination warns that truncating below roughly 1,024 dimensions degrades search quality when paired with binary compression. The team still wants the smallest dimensionality that respects that floor.
Which configuration fits?
Show answer and explanation
Answer:
text-embedding-3-large is trained with MRL, which "produces multiple vector representations at different levels of compression," and Azure AI Search exposes this through a truncationDimension property used together with scalar or binary quantization. Microsoft's own recommendation is to use "1,024 or higher for truncationDimension with binary quantization," warning that "a dimensionality of less than 1,000 degrades the quality of search results when using MRL and binary compression." 1,024 is the smallest value that respects that floor, giving the team the storage reduction they want without the quality loss they were warned about.
- A is wrong because 256 dimensions is technically a supported truncation point for text-embedding-3-large, but it sits well below the 1,024 floor Microsoft warns against for this exact quantization combination; it buys extra storage savings at the cost of degraded relevance. - C is wrong because it uses none of the compression the team is trying to achieve: the full 3,072-dimension vector is the largest, most expensive option to store, defeating the stated goal. - D is wrong because text-embedding-ada-002 predates MRL. Its 1,536-dimension output is fixed and can't be truncated, so it offers no path to the storage reduction the team wants.
Reference: Truncate dimensions with Matryoshka Representation Learning