Swapping pgvector for turbopuffer made search way faster and cheaper

The author explains how migrating an AI application from Postgres with pgvector to turbopuffer reduced p90 search times by 76% and significantly lowered database memory pressure and costs. The team initially encountered severe performance degradation on small accounts due to shared memory contention, and found that standard HNSW vector indexes worsened multi-tenant filtering. Although a binary quantization approach improved search latency, it failed to relieve database compute and memory costs until the embeddings were moved to a dedicated vector search solution. Open article →