Why I Use It
I run RAG applications and semantic search without adding another infrastructure piece. pgvector gives me cosine similarity right in Postgres, which I'm already running. Performance is fine for most workloads, and I don't have to sync data between a database and a vector store.
When I Recommend It to a Client
When their vector needs are moderate (millions of embeddings, not billions) and they want to avoid operating a separate vector database. Also when they already have Postgres in their stack.