While experimenting with personal, local RAG app setups, I kept having to (re)generate embeddings and really wanted precomputed embedding datasets that I could quickly pull and use in various environments.
I built a library to test out the idea: lance-bundle lets you package precomputed embedding vectors alongside the actual embedding model so that everything can be loaded from a single file for querying against the vectors; initial version uses LanceDB + ONNX for low dependency footprint and fast cold start to vector queries.
https://github.com/cloudkj/lance-bundle
As part of this, a few datasets that might be of interest to this audience have been precomputed as embedding vectors and hosted on a Hugging Face dataset hub and can be directly loaded and queried against:
https://huggingface.co/lance-bundle/datasets
With these datasets, you can simply load directly and run semantic queries to retrieve the nearest documents/embeddings:
from lance_bundle import load_dataset
bundle = load_dataset("lance-bundle/berkshire-hathaway-letters")
bundle.search("What does Warren Buffett think of passive index funds?")