Matryoshka Embeddings (MRL) What MRL Is Matryoshka Representation Learning trains an embedding so that the FIRST N dimensions are themselves a usable lower-dim embedding. Truncate the tail and renormalize — no retraining, no loss of most of the quality. Standard embedding (1536-d) truncated to 256-d: random quality loss (15-30%). MRL embedding truncated to 256-d: near-baseline quality (often within 2-5%). Models With Official MRL Support | Model | Native dims | Recommended truncations | |---|---|---| | OpenAI text-embedding-3-small | 1536 | 512, 768, 1024 | | OpenAI text-embedding-3-large | 3…