I'm not sure I understand. Isn't vector storage about storage and retrieval strictly, being a separate thing from the embedding generation itself? My mental mo

I'm not sure I understand. Isn't vector storage about storage and retrieval strictly, being a separate thing from the embedding generation itself? My mental model is that the job of vector storage is to just fetch the closest vectors to a target vector for some proximity metric, with the model (or method , more generally) that generates embeddings being its own distinct problem. You may later do things like passage re-ranking and maximal marginal relevance to refine it, but again, different task. I understand vector search against tons of other vectors is an approximation endeavor due to practical concerns, so it doesn't seem strange for vector store accuracy to be measured in terms of top-k retrieval. Post processing and the quality of the vectors themselves is a different thing to judge.