What is Milvus?
Milvus is a purpose-built vector database designed to store, index, and query high-dimensional embedding vectors at scale. It powers similarity search across billions of vectors with sub-second latency, making it the backbone of modern RAG pipelines, recommendation engines, and semantic search systems.
Core Concepts
Collections
Analogous to tables in SQL. A collection holds entities, each with a vector field and optional scalar metadata fields.
Partitions
Logical subdivisions within a collection for faster filtered queries and efficient data management.
Index Types
IVF_FLAT, HNSW, DiskANN — trade-offs between recall, speed, and memory. Choose based on your precision requirements.
Similarity Metrics
L2 (Euclidean), IP (Inner Product), COSINE — select the metric that matches your embedding model's training objective.
Architecture at a Glance
Milvus separates compute and storage. The proxy layer handles client connections, query nodes execute searches, data nodes manage persistence, and the index nodes build GPU-accelerated indexes. Object storage (MinIO or S3) holds the raw data, while etcd coordinates metadata across the cluster.
Meridian + Milvus
Meridian uses Milvus as its semantic memory layer. Every document chunk is embedded via our proprietary model, stored in a collection with hybrid scalar filtering, and retrieved at query time using HNSW for low-latency ANN search. The result: instant, context-aware answers grounded in your data.
Next Step: Dive into the Milvus Setup Recipe to spin up your first collection and run a similarity query in under five minutes.