Notes / Foundations
One changes what the model knows. The other changes how it behaves. Choosing between them is usually a category error.
The question is usually posed as a choice. It is not really one, because the two things do different jobs.
Retrieval changes what the model knows. Put relevant text in the prompt and the model answers from your data instead of its training.
Fine-tuning changes how the model behaves. Tone, output format, the shape of its reasoning.
If the refund policy changes on Tuesday, what has to happen?
With retrieval: you update a document. With fine-tuning: you assemble training data, run a job, evaluate it, and have a rollback story ready.
For facts that change, that is the whole argument.
Retrieval is auditable. Every answer in my pipeline returns the passages that produced it, with their scores. When a customer disputes an answer, I can show which chunk caused it.
A fine-tuned model cannot tell you why it said something. For anything touching money, policy, or a regulated domain, that traceability is not a nice-to-have.
They also compose. Fine-tune for format and tone, retrieve for facts.
Reaching for fine-tuning because retrieval quality is poor.
If the right passage is not being retrieved, fine-tuning does not fix it — you have taught the model to sound more confident about the same missing information. The fix is upstream: chunking, then hybrid retrieval, then reranking.
Fix retrieval before you consider training anything.