Retrieval or fine-tuning for this?

As of 12 August 2026: retrieval. If the question is about knowledge the model does not have — client documents, policy, product data — fine-tuning does not put it there reliably and retrieval does. Fine-tune when the problem is output format or style at high volume, or when the target is offline hardware that a 30B model does not fit. In both cases measure the stock model first.

Evidence: x-slm-edge-classifier, where a fine-tuned 7B ended up 2 points behind a stock 32B with a good prompt on the same classification eval (c-slm-edge-tuning-1); and g-fine-tune-domain-slm, where a domain fine-tune on banking classification never beat frontier once the eval was held constant. The pattern across both: the fine-tune closed a gap that had not been measured and turned out not to exist (c-slm-edge-tuning-2).

Caveat: 'almost always' has a known exception — very high-volume, narrow, format-bound tasks where a small fine-tuned model's per-token cost beats retrieval plus a larger model. That is a cost question, not a capability one, and the ledger answers it. Vendor fine-tuning APIs changed pricing in July; recheck before quoting.