Why OpenAI Embeddings Still Work Even After You Truncate Them

Why do OpenAI embeddings still work even after you truncate a large number of dimensions? In this video, we explore Matryoshka Representation Learning (MRL), a training technique that allows embeddings to remain useful even when you use only a prefix of the full vector. This makes it possible to trade off accuracy, memory usage, and retrieval latency at inference time. We first look at how embeddings are normally trained using contrastive learning, and why standard embedding models do not guarantee that truncated vectors will work well. Then we see how Matryoshka Representation Learning modifies the training loss to make smaller prefixes of the embedding independently useful. Finally, we look at results from the original MRL paper and experiments with modern embedding models to understand how truncation affects retrieval performance. This idea is especially useful for systems that rely on vector search, semantic retrieval, and RAG (Retrieval Augmented Generation). ⏱️ Timestamps: 00:00 Truncated OpenAI Embeddings Still Work 00:46 The Idea Behind Matryoshka Embeddings 02:03 How Embeddings Are Normally Trained (Contrastive Learning) 03:28 How Matryoshka Representation Learning Changes the Loss 05:16 Why Matryoshka Embeddings Are Useful for Vector Search & RAG 06:21 Results from the MRL Paper 📖 Resources: Matryoshka Representation Learning Paper - https://arxiv.org/pdf/2205.13147 Weaviate BlogPost - https://weaviate.io/blog/openais-matryoshka-embeddings-in-weaviate Weaviate Podcast on Matryoshka Representation Learning(with one of the authors) - https://www.youtube.com/watch?v=-0m2dZJ6zos 🔔 Subscribe : https://tinyurl.com/exai-channel-link 📌 Keywords: #openai #embeddings #vectorsearch #retrievalaugmentedgeneration