Concept
Matryoshka Embeddings
Embeddings trained so that prefix slices remain useful — choose any dimension at query time.
Definition
Matryoshka Representation Learning trains embeddings so the first k dimensions are individually meaningful for any k. Users can truncate the vector to balance accuracy and storage at runtime. OpenAI text-embedding-3 and Voyage support this.
Common use cases
- Storage savings
- Multi-tier search
- Latency tuning