Matryoshka Representation Embeddings Quiz

Answers and explanations appear as you take the quiz.

1. What is the core property of a Matryoshka embedding?
   • Trained prefixes can serve as representations at multiple lengths
   • Every dimension is equally important and order doesn't matter
   • It can only be used at one fixed length
   • It must be re-encoded for each target size
2. How is a Matryoshka model trained to achieve this?
   • By training a separate model for each size
   • With a loss summed over multiple nested prefix dimensionalities
   • By applying PCA after standard training
   • By randomly dropping dimensions at test time only
3. What do you do at inference to get a smaller embedding?
   • Retrain the model from scratch
   • Run a separate compression network
   • Truncate to the first k dimensions and renormalize
   • Average all the dimensions together
4. Which documented API lets a caller request fewer dimensions for text-embedding-3 output?
   • The dimensions parameter
   • A PCA flag in the Matryoshka paper
   • A FastText n-gram option
   • A GloVe context-window setting
5. What is a primary practical benefit of Matryoshka embeddings?
   • Fewer coordinates may reduce storage and comparison work; benchmark retrieval quality and cost
   • They eliminate the need for any vectors
   • They guarantee perfect accuracy at all sizes
   • They only work for images
6. In a Matryoshka embedding, where is the most discriminative information concentrated?
   • In the final, trailing dimensions
   • In the earliest, leading dimensions
   • Spread uniformly with no ordering
   • Only in the middle dimensions
7. Which dimensionality schedule is an example of nested Matryoshka prefixes in the guide?
   • Only one fixed size of 512 dimensions
   • A decreasing schedule from 2048 to 1 dimension with separate models
   • 8, 16, 32, and progressively larger prefixes up to 2048 dimensions
   • A sequence of unrelated sizes requiring new model weights each time
8. How does truncating a Matryoshka embedding differ from applying a separate PCA model?
   • It requires retraining a different encoder for every dimension
   • It replaces the original vector with a random projection
   • It can only be done while the model is being trained
   • A shorter prefix can be sliced and renormalized from the same trained embedding without a separate projection model