LLM Adaptation Strategy
A deep dive on a RAG prototype that is nearly good enough. The interviewer wants to see whether you reach for a technique or for the error analysis.
The questions in this round
A RAG prototype still gives inconsistent answers. When would you improve retrieval, change prompting, or fine-tune the model?
You look at fifty failures. In thirty of them the right passage was retrieved and the answer contradicted it anyway. Where does that leave fine-tuning?
My manager has budget for a fine-tune this quarter and would like to use it. What do you tell them?