The moment where RAG answered from the notes well but still could not behave the way you imagined matches what I have seen: retrieval adds knowledge but not style or task behaviour. That is where LoRA earns its keep, since format and reasoning patterns live in the weights, not the context window. How did you decide which failures were retrieval gaps versus ones only fine-tuning could close?
Kartik N V J K
The moment where RAG answered from the notes well but still could not behave the way you imagined matches what I have seen: retrieval adds knowledge but not style or task behaviour. That is where LoRA earns its keep, since format and reasoning patterns live in the weights, not the context window. How did you decide which failures were retrieval gaps versus ones only fine-tuning could close?