A school district is weighing two ways to make its GenAI assistant knowledgeable about district-specific curriculum content. One option adapts the model by fine-tuning it on the curriculum; the other pairs a general-purpose model with a retrieval document store that holds the current curriculum materials. Curriculum documents are revised every semester and staff need answers that trace back to the current source material. Which adaptation approach best fits this need?

Retrieval pairs a general-purpose model with a document store the model searches at answer time, so every response can pull from whatever curriculum materials are currently loaded, with the source document cited alongside the answer. Fine-tuning instead adjusts the model's underlying weights using a training pass over a fixed snapshot of content, which works well for teaching a model a stable writing style or domain vocabulary but leaves its knowledge frozen at whatever the training snapshot contained. Believing fine-tuning removes the need for a lookup step ignores that the same frozen-snapshot problem returns every semester, since each curriculum revision would require a brand-new training run to catch up. Believing fine-tuning removes the district's document-maintenance burden confuses the training process with content ownership; the district still has to prepare, review, and hand off the same curriculum materials either way. Assuming both approaches automatically track new revisions overlooks that only the retrieval store reflects a document the instant it is uploaded, while a fine-tuned model does not know about it until retrained. One scope caveat: fine-tuning can still be worth pursuing later for tone or format consistency, but that is a separate decision from keeping facts current. A useful operational check is asking whether last week's curriculum change already shows up in the assistant's answers without any retraining step.