Use a clean decision rule.
Prompting — the knowledge is already in the model and stable, or you just
need formatting/behavior; cheapest, try first
RAG — the answer depends on external, changing, or proprietary facts
that must be grounded and cited (a patient's report, current guidelines);
this is the genomics case
Fine-tuning — you need new behavior/style/format or to internalize a narrow
domain's patterns, and you have good training data; not for injecting fresh
facts (it's lossy and stale-prone)
They compose: fine-tune for domain tone, RAG for facts, prompt for control. For "explain my variant," RAG is correct because the facts are per-user and must be cited.