sapienzanlp/Minerva-7B-instruct-v1.0-GGUF · Which RAG embedding model to use with sapienzanlp/Minerva-7B-instruct-v1.0-GGUF

I deployed sapienzanlp/Minerva-7B-instruct-v1.0-GGUF on a virtual machine. It works, and the answers to my prompts sound fine.

As soon as I trigger the RAG system by uploading a document through the webui function, or add some knowledge, the model stops answering in a consistent way. Sometimes it simply replicates my question. Sometimes it answers by including the RAG prompt.

I assume I should set a different embedding model - the default one, all-MiniLM-L6-v2 seems optimised just for english.

Do you have any suggestion on a right model to choose that allow RAG with Minerva7b, or any other parameter I should take care ok?

Thanks