bge-m3

[Mimicvat]

https://huggingface.co/vonjack/bge-m3-gguf
from: ggerganov/llama.cpp#6007
I am looking for recommendations on a high-quality multilingual embedder that includes support for Portuguese. Anything better than https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2 would be nice.

[martinkozle]

I've been using https://huggingface.co/intfloat/multilingual-e5-large and it is pretty good.
I think https://huggingface.co/intfloat/multilingual-e5-large-instruct is newer and a little bit better.

[Mimicvat]

I've been using https://huggingface.co/intfloat/multilingual-e5-large and it is pretty good. I think https://huggingface.co/intfloat/multilingual-e5-large-instruct is newer and a little bit better.

I wanted to use this embedding too! How are you using it on ollama?

[martinkozle]

I am using Text Embeddings Inference from HuggingFace. But it has it's differences. Like having separate images for different hardware acceleration, or that the model can't be changed dynamically at runtime like with Ollama. I haven't tried using custom embedding models in Ollama yet, but I am interested when I have the time to give it a go. I don't know how the performance compares, I would believe that Text Embeddings Inference is faster?
I am interested to know what you manage to do.

[Mimicvat]

That's awesome! I wanted to create a simple RAG demo using AnythingLLM or another plug-and-play interface. However, I was looking for a good multilingual embedding to compare with the default one provided by AnythingLLM. I discovered that there are no state-of-the-art multilingual embeddings in ollama, even though it is integrated into every easy-to-use framework.