Milvus 2.6 Preview: 72% Memory Reduction Without Compromising Recall and 4x Faster Than Elasticsearch

What this post added

Milvus 2.6 introduces RaBitQ 1-bit quantization with adjustable Refine optimization (SQ4/SQ6/SQ8) for significant memory reduction and performance improvements. It also enhances full-text search with a revamped multi-language analyzer (Lindera, ICU, enhanced Jieba), adds Phrase Match with a 'slop' parameter for word order control, and introduces Time-Aware Decay Functions (exponential, Gaussian, linear) for relevance ranking based on document age. A new Function interface allows direct integration of embedding models (OpenAI, AWS Bedrock, Google Vertex AI, Hugging Face) for streamlined data ingestion and vectorization. Architecturally, a hot-cold tiered storage system is introduced for cost-effective scaling, and a new Streaming Node enables real-time vector processing with direct integration to streaming platforms.