AI Umum

Matryoshka Embedding Models: Producing Useful Embeddings of Various Dimensions


In the field of Natural Language Processing (NLP), embedding models play a crucial role in converting complex items like text, images, and audio into numerical representations that computers can understand and interpret. These embeddings, which are essentially fixed-size dense vectors, form the basis for various applications such as clustering, recommendation systems, and similarity searches. However, as these models become more sophisticated, the size of the embeddings they generate also increases, leading to efficiency challenges for downstream tasks.

Matryoshka Embeddings: A Novel Approach

To address this issue, researchers have introduced a new approach called Matryoshka Embeddings, which are models that produce useful embeddings of various dimensions. The name “Matryoshka” is inspired by the Russian nesting dolls that hold smaller dolls inside, reflecting the idea behind these models.

Matryoshka Embeddings are designed to generate embeddings that capture the most critical information in the embedding’s initial dimensions, allowing them to be truncated to smaller sizes without significant loss in performance. This feature enables variable-size embeddings that can be scaled according to storage requirements, processing speed demands, and performance trade-offs.

Applications of Matryoshka Embeddings

Matryoshka Embeddings have a wide range of applications. For instance, reduced embeddings can be used to quickly shortlist candidates before performing more computationally intensive analysis using the full embeddings, as in the case of nearest neighbor search shortlisting. This flexibility in controlling the embeddings’ size without compromising accuracy offers significant advantages in terms of efficiency and scalability.

Training Matryoshka Embedding Models

Compared to conventional models, Matryoshka Embedding models require a more nuanced approach during training. The process involves evaluating the quality of embeddings at different reduced sizes as well as at their full size. A specialized loss function that assesses the embeddings at multiple dimensions helps the model prioritize the most important information in the first dimensions.

Implementation and Practical Use

Frameworks like Sentence Transformers support Matryoshka’s models, making it relatively straightforward to implement this approach in practice. These models can be trained with minimal overhead by employing a Matryoshka-specific loss function across various truncations of the embeddings.

When using Matryoshka Embeddings in practice, embeddings are generated as usual, but they can be optionally truncated to the desired size. This technique reduces the computational overhead, significantly improving the efficiency of downstream tasks without hindering the embedding formation process.

Effectiveness and Demonstration

The effectiveness of Matryoshka Embeddings has been demonstrated using two models trained on the AllNLI dataset. The Matryoshka model outperforms a regular model across various metrics. Notably, the Matryoshka model retains approximately 98% of its performance even when reduced to just over 8% of its original size, showcasing its effectiveness and offering the potential for substantial processing and storage time savings.

The team has also showcased the capabilities of Matryoshka Embeddings through an interactive demo that allows users to dynamically change an embedding model’s output dimensions and observe how it affects retrieval performance. This practical demonstration not only highlights the adaptability of Matryoshka Embeddings but also underscores their potential to transform the effectiveness of embedding-based applications.


Matryoshka Embeddings have emerged as a powerful solution to the growing challenge of maintaining the efficiency of embedding models as they scale in size and complexity. These models open up new avenues for optimizing NLP applications across various domains by enabling the dynamic scaling of embedding sizes without compromising accuracy.