The authors fused SuperKMeans with e-means into SuperEMeans, achieving roughly 3x faster clustering of high-dimensional vectors with no quality trade-off. The work targets the clustering bottleneck in building large-scale vector indexes such as IVF and ScANN, with open-source code and reproducibility scripts on GitHub.
Vector Embedding
向量嵌入
High-dimensional dense vector representations of text, images and other objects, foundational to vector search and RAG.
Curated Intelligence
Perplexity migrated its core search service from Amazon DynamoDB to CobbleDB, a custom Rust-based distributed key-value store, to address latency and cost bottlenecks in high-throughput LLM retrieval. By decoupling persistent storage from the hot data layer, the team achieved a fivefold reduction in batch read latency and cut storage costs by at least 20%. The system was built in two months by two engineers working with AI coding agents and is planned for open-source release.
Google released the open-source multimodal embedding model EmbeddingGemma 2, which outperforms rivals twice its size on benchmarks despite having only 740 million parameters. The model supports local execution with minimal RAM usage (191 MB), significantly reducing vector database storage requirements and improving retrieval efficiency.
NVIDIA's technical team released a comprehensive open-source workflow enabling developers to build domain-specific embedding models within a day. Integrating Nemotron, NeMo Data Designer, and NeMo Automodel, the solution addresses the semantic limitations of general-purpose embeddings in vertical domains (e.g., contracts, manufacturing logs), significantly enhancing RAG retrieval accuracy.
OpenAI released contrastive pre-trained embeddings for text and code to enhance semantic retrieval and understanding. The study leverages large-scale unsupervised training to optimize vector representations, supporting applications like RAG and code search.