The obvious question would be how does it compare to dedicated vector databases (qdrant, pinecone, chroma, etc..). What are the tradeoffs?