What should this product use: rules, RAG, fine-tuning or a task-specific SLM?

Choose the narrowest honest description of the decision the product must make.

This distinguishes knowledge retrieval from behaviour adaptation.

Use the field condition, not the laboratory or headquarters condition.

Do not confuse user preference with a system-level control deadline.

Consider customer commitments, classification, trade secrets and deployment policy.

The deployed hardware envelope can eliminate otherwise attractive options.

Fine-tuning is for repeatable behaviour—not for keeping facts current.

Select what is actually governed and usable, not what might be collected later.

A language model must never silently become the deterministic safety controller.

Architecture commitment should follow evidence, not enthusiasm.

Customisation creates evaluation, update, rollback and incident obligations.