Three levers, three different jobs
Prompting changes instructions. Retrieval changes what facts are in front of the model at answer time. Fine-tuning changes the model's weights, and with them its default behaviour and format. Most production problems are one of the three, and teams routinely reach for the wrong one.
Why it is like this
They act at different points: prompting at the instruction layer, retrieval at the evidence layer, fine-tuning at the behaviour layer. Treating them as interchangeable effort is what produces expensive projects that fix nothing.
Not to be confused with
A ladder of sophistication where fine-tuning is the grown-up option. Order by cost and reversibility: prompt first because it is instant, retrieval next because it is auditable, fine-tune last because it is slow, needs data, and has to be redone when the base model changes.
In practice
"It does not know our products" is retrieval. "It will not answer in our format" is prompting, then fine-tuning. "It is wrong about our products" is retrieval quality, not model choice.
Source: Lewis et al., Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (arXiv:2005.11401) — §1. Read on .
The OWASP Top 10 for LLM Applications is © the OWASP Foundation and published under a Creative Commons licence; NIST AI 600-1 is a US Government work in the public domain. These cards are independent commentary written in our own words. ytyk.org is not affiliated with or endorsed by the OWASP Foundation or NIST.
Predict Check yourself
A support bot gives outdated answers about your pricing. Which lever?
Show the answer
Retrieval — the facts in front of the model at answer time are stale; fine-tuning would bake today's prices into weights that cannot be updated when they change.
Facts that change belong in retrieval, where they can change too. This is the single most common misallocation of effort in GenAI projects.