🪆 Introduction to Matryoshka Embedding Models

Hugging Face Blog · 2 years ago

Matryoshka embedding models allow embeddings to be truncated to smaller dimensions while retaining performance, enabling storage and speed tradeoffs for tasks like retrieval and search. In experiments comparing a Matryoshka model to a standard model on STSBenchmark, the Matryoshka model preserved 98.37% of performance at 8.3% of full embedding size, versus 96.46% for the standard model. This approach makes it practical to deploy embedding systems across different storage budgets and processing speeds without significant accuracy loss.