tazarov
commented
2 months ago @h3clikejava, thanks for raising this. Let's start with some background:
- Default EF uses onnx runtime + mean pooling, resulting in 384-dimensional embeddings
- ST does everything out of the box and again produces 384-dimensional embeddings
Under normal circumstances, you should not have trouble swapping between the two, as Chroma will accept queries using 384-dimensional embeddings even though there are slight differences in the output embeddings (in the order of 1-e4/1-e5 range).
That said, when you say it won't work, do you mean you get an error? Can you share the error?
jeffchuber
What happened?
I save some embeddings by default like this: collection = client.get_or_create_collection(name=db_name)
Then, i can fetch data by DefaultEmbeddingFunction() like: emb_fn = embedding_functions.DefaultEmbeddingFunction() collection = client.get_or_create_collection(name=db_name, embedding_function=emb_fn) # It's work
But i can't fetch data by all-MiniLM-L6-v2 like: emb_fn = embedding_functions.SentenceTransformerEmbeddingFunction(model_name="all-MiniLM-L6-v2") collection = client.get_or_create_collection(name=db_name, embedding_function=emb_fn) # It's not work
What's the different between DefaultEmbeddingFunction and SentenceTransformerEmbeddingFunction(model_name="all-MiniLM-L6-v2")?
Versions
ChromaDB V0.5.3, python v3.10.11, MacOS 15.0 Beta (24A5327a)
Relevant log output
No response