What are the different vector database options for RAG and how do they compare?

Vector database comparison for RAG: Pinecone offers fully managed service with excellent scaling, metadata filtering, and hybrid search but is cloud-only and costs scale with vectors stored. Qdrant provides open-source flexibility with rich filtering, quantization, and multi-tenancy support, deployable on-premise. Weaviate features built-in vectorization modules, GraphQL API, and hybrid search, good for prototyping with its generative modules. Milvus excels at massive scale (billions of vectors) with GPU-accelerated indexing and multi-vector search. Chroma is lightweight and developer-friendly, ideal for prototyping and small-scale deployments. pgvector integrates with PostgreSQL, enabling vector search alongside relational data in one database, excellent for applications already using PostgreSQL. Selection criteria: scale (millions vs billions of vectors), latency requirements (Pinecone and Qdrant lead at sub-10ms), filtering complexity (Qdrant and Weaviate have the richest filtering), operational overhead (managed vs self-hosted), cost (pgvector and Chroma are free, managed services charge per vector/query), and existing infrastructure (pgvector wins if you already use PostgreSQL/Supabase). For most RAG applications, pgvector with Supabase provides the best balance of simplicity, cost, and capability.