AI decision
RAG vs Fine-Tuning
Short answer: use RAG when the answers live in documents you can retrieve. Use prompting when the job is a short, owned FAQ. Fine-tune only after a hosted model plus RAG still fails a real eval set.
What each is for
RAG fetches chunks at request time. Fine-tuning changes model weights so it “sounds like” your domain. Fine-tuning does not give the model last week’s PDF. RAG does not teach tone as well as a small fine-tune — and you usually do not need tone in v1.
MVP default
Hosted API → better prompt → RAG if there are files → eval set of 20 questions. Fine-tune is a later experiment, not a launch blocker.
FAQ
Can I do both?
Yes, later. Doing both in week one is how budgets disappear.