LocalLLaMA

RWKV v5 7b, its only half trained rn, but the model surpasses Mistral on all multilingual benchmarks, cause the is meant to be multilingual.

1. you’d probably want to embed the notes and then use cosine similarity to find similar notes given your input query (“please write a short sentence describing the following facts…”, use the facts in the cosine similarity search ).

2. Then pass the similar notes into a llm with a instruction like “please write a short sentence describing the following facts using the notes”

I don’t know how well embeddings work for italian so you may want to translate them to english and keep them in pairs (italian version, english version) then use the english for the cosine similarity search ( step 1. ) and the italian version for the summarization ( step 2. )

You can try taking a picture of the notes and having a multimodal model try and read and extracts its text. You can either use chatgpt4(paid but probably more accurate) or run llama.cpp llava multimodal function with a llava model locally(free but might hallucinate).

​

Maybe scanning your notes into PDF format and trying a RAG approach might yield some results too. You can upload the PDF to GPT/Claude or run a local RAG project like h2oGPT or privategpt and see how well they can transcribe your notes.