Knowledge Cutoff
The date the model’s knowledge stops.
The point after which the model has no training data. Anything newer is unknown to it unless it is retrieved or pasted in.
Cutoffs are a common source of confidently wrong answers about recent events, library versions and your own product. If freshness matters, retrieval is the mechanism, not the model.
Two practical consequences. Anything time-sensitive must come from retrieval or a tool, because a model confident about last week is guessing. And if you do not tell the user the cutoff, they will over-trust answers about recent events — the systems where this failed publicly were missing a date in the interface, not a better model.
A date in the UI, not just the prompt
A user asks about pricing announced after the model’s training cutoff.
Tell the model its cutoff so it can hedge, and show a "last updated" date on the answer. The wrong-answer risk is reduced by retrieval, but the trust risk is only fixed by telling the reader what the system knows and when.
- guideIntroduction to large language models ↗
A plain-language official introduction to models, tokens and context. Teaching material rather than origin research — good for the definition, not for a claim about the frontier.
- docsWeb search ↗
Distinguishes a quick lookup from agentic search, where a reasoning model searches, judges the results and searches again. Also covers source annotations.
Parametric Knowledge
What the model stores inside its weights.
RAG · Retrieval-Augmented Generation
Look things up, then put what you found in the prompt.
Hallucination
Fluent output that is not true.
Contextual Knowledge
What the model knows from the prompt, right now.