Benchmarked 5 vector databases on 1M vectors at 1536 dims. Latency, recall, and cost

Latency (p50):

• Qdrant: \~3ms
• Milvus: \~4ms
• pgvector HNSW: \~5ms
• Weaviate: \~8ms
• Pinecone Serverless: \~12ms

pgvector is way faster than people think. The "Postgres is slow for vectors" thing is from the IVFFlat era. HNSW changed that

Recall:

• Qdrant: 99.2%
• Milvus: 99.0%
• pgvector: 98.5%
• Pinecone: \~95% (can't tune HNSW params)

Monthly cost (1M vectors):

• pgvector on Neon: $30-150
• Pinecone Serverless: $50-80
• Qdrant Cloud: $65-102
• pgvector on RDS: $260

We use pgvector on Neon in production. Replaced Pinecone and RDS, latency dropped from 200ms to 80ms