Decision tree

RAG, fine-tune, prompt — or hybrid.

The single most common LLM-product mistake is reaching for fine-tuning when the problem is knowledge, or for RAG when the problem is behaviour. The two problems look similar; the right shape is opposite. This tree walks the fork.

4 questions max7 end-states~3-5 minShareable via URL hash

If JavaScript is disabled — the questions in this tree

  1. Does your use-case depend primarily on knowledge or on behaviour?
  2. How often does the knowledge change?
  3. How many examples of the desired behaviour can you reliably curate?
  4. Is the behaviour requirement strict (e.g. always-JSON, always-cite, always-escalate-on-X)?

Re-enable JavaScript to step through interactively. Or jump straight to the related artefacts: diagnostics · reference architectures · writing.

Also on this site