RAPIDS cuVS

NVIDIA's generalist embedding model achieving record 69.32 score on MTEB benchmark. Fine-tuned from Llama architecture with improved techniques for training LLMs as embedding models.

An open-source benchmarking tool from Zilliz for comparing vector database performance and cost-effectiveness. Provides an intuitive visual interface to reproduce results and test new systems with standardized metrics.

NVIDIA RAPIDS cuVS is a GPU-accelerated library for vector search and clustering with CUDA-optimized HNSW, IVF, CAGRA, and PQ implementations. Supports batch queries for high QPS, suited for large-scale similarity search in GenAI apps. Achieves up to 12x faster indexing and lower latency vs CPU-only alternatives like FAISS CPU.

Comprehensive benchmark dataset comparing 10 vector databases across 19 fields including query latency (p50/p99), throughput, scalability limits, features like hybrid search and ACID compliance, SDK support, and managed pricing. Tested with 1M vectors at 1536 dimensions for RAG and AI search applications. Key highlights include Qdrant for lowest latency, Pinecone for managed scalability, and pgvector for ACID transactions.

Open-source vector database benchmarking tool testing databases across production-critical scenarios including static collection, filtering, and streaming cases with modern embedding model datasets.

Open-source comparative benchmarks evaluating vector search performance of engines like Qdrant, Elasticsearch, Milvus, Redis, and Weaviate. Covers single-node upload/search, filtered search across various datasets and configurations, focusing on RPS, latency, precision, and indexing time using affordable hardware.