Pinecone

Managed serverless vector database with a storage-compute separation architecture built directly on Amazon S3 (and equivalent object stores on GCP/Azure). Vectors live in immutable **slab** files on S3; queries run against a stateless query-executor fleet that caches slabs on local SSDs. The platform exposes a hosted API + SDK with extras including hosted embedding/reranking models (Pinecone Inference), production chat-agent scaffolding (Pinecone Assistant), and dedicated read-only nodes for read-heavy workloads.

Traditional vector databases force operators to provision RAM-sized indexes — billion-vector workloads need ~3TB of memory for a 768-dim float32 HNSW graph alone, which is operationally prohibitive. Pinecone's serverless architecture decouples storage (cheap, in S3) from query compute (elastic, pay-per-query), letting customers run trillion-scale vector indexes without capacity planning. The 2024 launch reset the cost curve for hosted vector search; the 2025-2026 evolution added BYOC mode so the data plane can run in customer-owned AWS/GCP/Azure accounts.

Production RAG pipelines for chatbots and agentic AI, semantic search over millions-to-billions of documents/embeddings, multi-tenant SaaS products (each tenant gets its own namespace), AI-assistant backends where freshness matters (Pinecone Assistant), and bring-your-own-cloud deployments where compliance forbids cross-cloud data movement.