Retrieval-Augmented Generation

knowledge cutoff

An LLM's built-in knowledge stops at a date — the point when the snapshot of training data was taken. Ask about anything that happened after that knowledge cutoff and the model is essentially guessing, because those events were never in what it learned. This is why a model can be cheerfully unaware of a product released last month or a result announced last week.

The cutoff is fixed at training time and no amount of clever prompting moves it. The only real ways past it are retraining the model, fine-tuning it on newer data, or — by far the cheapest and most common — RAG, which feeds current documents in at the moment of the question. Worse, a model usually will not warn you that you have wandered past its cutoff; it answers stale facts in the same confident tone as fresh ones, so out-of-date information can look exactly like reliable information.

Also called
training cutoff