AI-300 Practice Question 30: Vector Storage Optimization

Microsoft's vector optimization guidance lists quantization first: it "compresses native float32 or float16 embeddings to int8 (scalar) or byte (binary)" and "reduces storage in memory and on disk". It then says MRL "supplements scalar and binary quantization. When you use either quantization method, you can specify a `truncationDimension` property on your vector fields to reduce the dimensionality of text embeddings", supported because the text-embedding-3 models "are retrained on the Matryoshka Representation Learning (MRL) technique that produces multiple vector representations at different levels of compression". Microsoft adds that "using binary quantization with MRL provides the maximum vector index size reduction". Both levers apply to the model already in use, and truncated dimensions "can be set only when scalar or binary quantization is configured", which is why the answer pairs them (option A).

- B is wrong because exhaustive KNN changes how the index is *searched*, not how large the vectors are. It scans every vector on every query, so at 12 million chunks it raises query latency and compute cost, the opposite of the goal. - C is wrong because it removes semantic matching altogether rather than optimizing it. Relevance would not stay close to today's for paraphrased or conceptual questions, which is the constraint the team set. - D is wrong because chunk overlap exists to preserve continuity across chunk boundaries, so cutting it degrades retrieval quality. It also barely moves the number of chunks, and it leaves the per-vector size, the thing that actually dominates the bill, completely unchanged.

Reference: Truncate dimensions using MRL compression - Azure AI Search