Matryoshka Representation Learning
An embedding training technique that enables model-generated vectors to be truncatable, such that prefix sub-vectors of the same high-dimensional vector retain most semantic information. This allows dynamic adjustment of vector dimensionality according to resource constraints, making it possible to use a single trained embedding at multiple granularities without retraining.
Timeline (last 90 days)
EmbeddingGemma2 employs Matryoshka Representation Learning, encoding the most important information at the front of the vector to support vector truncation.
Google has also adopted Matryoshka Representation Learning technology in models such as text-embedding-004.
EmbeddingGemma 2 adopts Matryoshka Representation Learning, allowing dimensions of 256, 128, or fewer to still be used as standalone embeddings.
All Facts (3)
EmbeddingGemma2 employs Matryoshka Representation Learning, encoding the most important information at the front of the vector to support vector truncation.
50%UnverifiedGoogle has also adopted Matryoshka Representation Learning technology in models such as text-embedding-004.
50%UnverifiedEmbeddingGemma 2 adopts Matryoshka Representation Learning, allowing dimensions of 256, 128, or fewer to still be used as standalone embeddings.
50%