Strong take on keeping embeddings in PostgreSQL instead of a separate vector store. We made the same decision for a document search system — single-transaction writes between documents and their embeddings eliminated an entire class of consistency bugs we had with a Pinecone setup. The tradeoff showed up at around 2M vectors where pgvector query latency started climbing, but HNSW indexing brought it back under 50ms.
This is exactly the kind of real-world data point the article needed — thank you for sharing it. The single-transaction write eliminating consistency bugs is the argument I find hardest to quantify on paper but most compelling in practice. Once you've debugged a system where documents and embeddings can drift, the operational simplicity of keeping them in one transaction becomes very difficult to give up.
The 2M vector threshold with HNSW bringing latency back under 50ms is genuinely useful — that's a more precise signal than most teams have when they're deciding whether to stay on pgvector or move to a dedicated store.
klement Gunndu
Agentic AI Wizard
Strong take on keeping embeddings in PostgreSQL instead of a separate vector store. We made the same decision for a document search system — single-transaction writes between documents and their embeddings eliminated an entire class of consistency bugs we had with a Pinecone setup. The tradeoff showed up at around 2M vectors where pgvector query latency started climbing, but HNSW indexing brought it back under 50ms.