KnowledgeBase requires recomputation of embeddings upon each run, we could have serialisation methods, or a method to load it with embeddings.

[davidberenstein1957]

[GTimothee]

Will do!

[davidberenstein1957]

Lovely. Thanks @GTimothee. I think the issues holds 2 sub-issues. Potentially, allowing to re-use, start with pre-computed embeddings, and adding serialisation to the KnowledgeBase (saving and loading locally, and to the Hub). We can copy the logic and behavior as much as possible from the TestDataset you worked on :)

[GTimothee]

fyi working on it ;) (https://huggingface.co/datasets/GTimothee/my-knowledge-base)