Start with pgvector: Why You'll Outgrow It Faster Than You Think - Qdrant

What this post added

This post provides a detailed comparison of pgvector and dedicated vector databases, outlining specific technical limitations of pgvector related to dataset size (1M vectors), metadata filtering (lack of pre-filtering), hybrid search (missing BM25), and architectural coupling. It contrasts these with the advantages of dedicated stores, such as efficient filtering, native hybrid search (reciprocal rank fusion), scalability beyond 10M vectors, and decoupled architecture. It also introduces Qdrant's native multivector support and filterable HNSW as solutions to pgvector's shortcomings.