Hello,
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?")
Looking to share to see if anyone actually finds it useful, and to gather feedback on whether it makes sense for the local-first AI enthusiasts. Let me know what you think!Comments URL: https://news.ycombinator.com/item?id=49259385
Points: 1