Post
Highlighted: the lines this signal was extracted from
We run an image and semantic text search API. The first version used Pinecone for vectors and RDS for everything else and the hardest part was keeping the two in sync and performant About a year ago we moved everything into Neon with pgvector (Neon wrote up the migration here: Neon blog post). End to end search went from about 200ms to 80ms, mostly from cutting a network hop. The vector query itself takes 3-4ms (Tested on 10M records) Since our backend is on AWS and Neon uses AWS as well, we cut the network latency down to couple ms Happy to answer anything about the setup if you're thinking about doing the same