Vector databases are the backbone of semantic search, RAG, and AI-powered recommendation systems. They store embeddings (vector representations of text, images, or audio) and enable fast similarity search — finding the closest vectors to a query. In 2026, the vector database market has consolidated around a few clear leaders. This comparison covers performance, scalability, and pricing for production deployments.

Quick Comparison

FeaturePineconeWeaviateQdrantMilvuspgvector
TypeManaged vector DBVector DB + knowledge storeVector search engineDistributed vector DBPostgreSQL extension
Open SourceNo (SaaS only)Yes (BSD-3)Yes (Apache 2.0)Yes (Apache 2.0)Yes (PostgreSQL License)
Self-HostedNoYes (Docker, K8s)Yes (Docker, K8s)Yes (Docker, K8s, distributed)Yes (any PostgreSQL)
Index TypesProprietary (serverless)HNSW, flat, IVFHNSW, quantization (scalar/binary)11 index types (HNSW, IVF, DiskANN, etc.)IVFFlat, HNSW
Metadata FilteringYes (metadata filtering)Yes (GraphQL + vector hybrid)Yes (payload filtering, rich query DSL)Yes (scalar filtering, expressions)Yes (SQL WHERE + ORDER BY distance)
Hybrid Search (dense + sparse)No (dense only)Yes (BM25 + vector)No (dense only, sparse via plugin)Yes (hybrid search)Yes (tsvector + pgvector)
Max Vectors (single node)Unlimited (serverless)~1B (with quantization)~1B~10B (distributed)~10M (practical), ~100M (with tuning)
Query Speed (1M vectors)5-20ms5-15ms2-10ms5-20ms10-50ms
Pricing$0.10/GB/month + $0.012/1M queriesFree (OSS), Cloud from $0.12/GB/moFree (OSS), Cloud from $0.15/hrFree (OSS), Cloud from $0.10/hrFree (PostgreSQL extension)

When to Choose Each Database

Pinecone — Best for: Teams that want zero ops and predictable pricing. Pinecone is the simplest to start with — create an index, insert vectors, query. No infrastructure to manage. Weak spot: SaaS-only (no self-hosting); dense vectors only (no sparse/hybrid); can get expensive at scale.

Weaviate — Best for: Teams that need hybrid search (dense + sparse) and want a knowledge graph feel. Weaviate stores both vectors and the original objects, supports GraphQL, and has the best hybrid search of any vector DB. Weak spot: More complex to configure than Pinecone; JVM-based (higher memory usage).

Qdrant — Best for: Performance-focused teams that want the fastest queries with rich filtering. Qdrant is written in Rust for maximum performance and has the best filtering DSL. Weak spot: Smaller managed cloud offering; less mature than Pinecone/Weaviate.

Milvus — Best for: Teams operating at massive scale (1B+ vectors) that need a distributed vector database. Milvus has the most index types and the best documentation for large-scale deployments. Weak spot: Most complex to operate; overkill for <10M vectors; distributed architecture requires multiple components.

pgvector — Best for: Teams that already use PostgreSQL and want to add vector search without a new database. pgvector is a PostgreSQL extension — same SQL you know, same backups, same infrastructure. Weak spot: Performance degrades significantly above ~10M vectors; HNSW index builds are slow; fewer index tuning knobs than dedicated vector DBs.

Decision Matrix

ScenarioBest Vector DBWhy
Start quickly, zero ops, predictable costPineconeEasiest to start, serverless, no infrastructure
Hybrid search (dense + sparse/keyword)WeaviateBest hybrid search, GraphQL-native
Performance + rich filteringQdrantFastest queries, best filtering DSL, Rust
Massive scale (1B+ vectors)MilvusMost scalable, most index types
Already use PostgreSQL, <10M vectorspgvectorZero new infrastructure, same SQL, same backups
Open source, self-hosted, moderate scaleQdrant or WeaviateBoth have strong self-hosted offerings

Bottom line: Start with pgvector if you already use PostgreSQL — it handles up to ~10M vectors well and eliminates the operational burden of a separate database. Move to Qdrant or Weaviate when you outgrow pgvector (performance, scale, or need advanced features like hybrid search). Pinecone is the easiest paid option — use it if you want zero ops and predictable pricing. Milvus is the pick for 1B+ vector scale. See also: RAG Best Practices and Embedding Models Comparison.