If the model is wrong because it cannot see your documents, fine-tuning will not help. Retrieval-augmented generation (RAG) is the default: ingest the corpus, preserve permissions, and cite sources.
Fine-tuning earns its keep when the failure is style, structure, or a long tail of domain forms the base model keeps mangling even with the right chunks in context. We only train when a held-out eval beats the RAG-only baseline by a margin you agree in advance.
The artefact you should own is the dataset and the eval, not a vendor’s unnamed “custom model.” That is what lets you change providers later without starting from zero.
