Makes a lot of sense. It really depends on the dataset and model you use for how the data is embedded within the space of all embeddings. If data is very similar you may find that you have to have a lot of candidates because internally we are having to explore a significantly larger portion of the underlying HNSW graph. 1000 candidates seems highish to me to get the best top 10 but it's definitely not crazy and this to me is a state of the current research problem where it's difficult to ascertain whether a model will generate good embeddings on your data without being an expert in the space. You might just try a few other models and see if you get less expensive results (better top 10 without so many candidates).