My understanding is that if you want to eg load your company employee handbook (~20 page PDF) into a model you're likely to get much better results with RAG than with fine-tuning – can you fine tune such that the model reliable answers questions about something that small?

→ View original post on X — @simonw,