AFAIK what is needed from that dict is a

Yea thats right. Is the question just about how to persist that data? Or how to build that dict?

@Logan M yes, I might need to persist that. To be able to create that dict.

is there any construct currently in llama index or do I need to build it to persist ?

memory wise I think it can be a challenge. I imagine a callable connected to some persistance to be able to scale. am I missing something?

Not currently -- persistance is up to you

Yea thats a good point. I wonder if a class that implements __getitem__ that calls to like redis or something, instead of an actual dict, would work well?

yeah, AFAIK what needs to work is this line :

node = self._node_dict.get(query_id, None)

so it might need a get function. AFAIK __getitem__ will only make [] this work

"something" that implements protocol with a get will do it.

@Logan M is there any functionality to save nodes ?

AFAIK you can use VectorStoreIndex to save the nodes and persist them. Later you can use load_index_from_storage to load from the persist_dir and then get nodes and build the node dict required for the RecursiveRetriever. Here is a github example : https://github.com/run-llama/llama_index/discussions/8930

the docstore is made to save and hold nodes

The vector store holds nodes as well, but that kind of assumes you want them embedded

ahh correct

kk, thank you.