Milvus 2.4 Unveils Game-Changing Features for Enhanced Vector Search
By Zilliz on Mar 20, 2024
Redwood City, CA—March 20, 2024— Zilliz, a trailblazer in vector database technology, is proud to announce the release of open source Milvus 2.4, introducing groundbreaking features designed to revolutionize vector search capabilities. With a focus on improving retrieval quality, developer efficiency, and resource optimization, Milvus 2.4 represents a significant leap forward in vector database technology.
Multi-vector / Multimodal Support
A new feature of Milvus 2.4 is its Multi-vector support, offering an end-to-end infrastructure solution for multi-vector search and reranker functionalities. This feature enhances retrieval quality and developer efficiency by reducing barriers associated with working with multi-modal vectors and reranking algorithms. Multi-vector support introduces two key components:
Storage and Query of Multiple Vectors: Milvus now enables storing and querying of multiple vectors for a single entity within a collection, providing a more natural way to organize data.
Built-in Reranking Algorithms: Developers can leverage prebuilt reranking algorithms in Milvus to build and optimize their reranking algorithms, further enhancing retrieval performance.
Grouping Search
Milvus 2.4 introduces a Grouping Search operation to enhance resource efficiency and developer productivity by ensuring users receive relevant and meaningful search results. By aggregating and organizing based on specified criteria, Milvus enables users to retrieve top-level entities (documents, audio, or image files) instead of individual chunks, enhancing the diversity, usability and accuracy of search outcomes and facilitating more effective data analysis and decision-making processes.
Sparse Embeddings
Milvus 2.4 introduces sparse vector support, facilitating efficient ANN search over sparse embeddings generated by models like SPLADEv2 or statistical models like BM25. Sparse embedding search considers semantic similarity, while keyword or inverted index only accounts for word frequency. This enhances the accuracy of text search using hybrid search methodologies, providing users with more nuanced and relevant search results.
Regular Expression & Inverted Index
Milvus 2.4 introduces powerful features for enhanced data filtering and performance optimization. With the addition of Regular Expression and Inverted Index capabilities, users experience improved substring matching during metadata filtering and benefit from significant performance gains, with up to a 30x increase in performance for filtering scalar data types.
First to Market with GPU Indexing / CAGRA
Undoubtedly, the most anticipated feature of Milvus 2.4 is the introduction of the groundbreaking GPU Indexing capability, which leverages NVIDIA's cutting-edge CUDA-Accelerated Graph Index for Vector Retrieval (CAGRA). This feature revolutionizes GPU-based vector search, surpassing traditional CPU-based indexes like HNSW, even with small batch sizes.
With remarkable performance gains, particularly under large data sets, Milvus 2.4 also supports brute force search for CAGRA implementations for situations prioritizing recall without sacrificing performance.
"For developers leveraging vector databases, our focus centers on three essential factors: user experience, accuracy, and performance. With Milvus 2.4, we've prioritized these critical aspects," said Charles Xie, Founder and CEO of Zilliz. "Features such as CAGRA support for enhanced performance, Multi-vector for improved retrieval quality and developer efficiency, and Regular Expression and Inverted Index capabilities highlight our commitment to driving innovation while prioritizing user experience."
Milvus 2.4 is now available for download. Explore the latest features and enhancements on the Milvus website.
About Zilliz
Zilliz is a leading vector database company, founded by the engineers who created Milvus, the world's most widely-adopted open source vector database. Zilliz's next-generation database technologies help organizations rapidly create AI/ML applications and unlock the potential of unstructured data. By simplifying complex data infrastructure management, Zilliz is committed to bringing the power of AI to every corporation, organization, and individual.
Headquartered in Redwood Shores, CA, Zilliz is backed by prestigious investors, including Aramco's Prosperity7 Ventures, Temasek's Pavilion Capital, Hillhouse Capital, 5Y Capital, Yunqi Partners, Trustbridge Partners, and others. Zilliz's technologies and products help over 5,000 organizations worldwide easily create AI applications in various use cases. Learn more at zilliz.com or follow @zilliz_universe.
share
Media assets
Official Zilliz family logos and assets for download.
Contact us
For media and analyst inquiries.