Truncate Matryoshka embeddings to 256 dimensions — top MRL models lose under 1% recall

Milvus·medium signal

2026 embedding benchmarks confirm Matryoshka-trained models can be truncated hard with minimal recall loss: Voyage and Jina v4 lose under 1% at 256 dimensions. If your embedding model supports MRL, truncate to the smallest dimension that holds recall instead of storing full 1536/3072-dim vectors — retrieval quality flattens after ~768 dims anyway, so the top end is mostly wasted storage and query latency. This is a config change, not a re-embedding project.