I'm curious that, for RecursiveRetriever or BM25Retriever or other retriever that requires to store the nodes as attribute, are they suitable for use cases where large amount of documents exist? Would that consume too much memory?

I see that VectorStoreIndex.docstore.docs should do the work

Hi, after building VectorStoreIndex, is there a way to get the nodes? I'm parsing a document into hierarchical nodes, where sub nodes are IndexNode referencing the root node. Using RecursiveRetriever, I have to provide either node_dict or retriever_dict or query_engine_dict. I assume providing retriever_dict or query_engine_dict requires building an index for each root node, which could be wasteful. So I prefer node_dict. But if the VectorStoreIndex is loaded from some external storage, how could I get those nodes back (if possible)