Skip to content
Search
TechniquesGlossary

Knowledge Cutoff

A large language model learns from text gathered up to a certain date. Anything after that, such as a new law, a price change or last week's news, isn't in its weights. Ask about it and the model may say it doesn't know, or it may answer confidently with outdated information. Providers publish a training data cutoff for each model; Anthropic's model overview, for example, lists one per model. Check the provider's documentation rather than asking the model, which may not know its own cutoff reliably.

The fix is to bring current information to the model. Paste the relevant document, connect retrieval (RAG) or a web search tool, and tell the model to answer from those sources (grounding). Put today's date in the prompt when timing matters, and ask the model to say when its answer may be out of date. That reduces confident wrong answers, a common form of hallucination.

Example

Today's date is {today}.
The tax rules below are current as of that date. Answer only from them.
If the question depends on anything not covered here, say so and don't rely on what you remember from training.

<rules>
{paste the current rules}
</rules>

Question: {question}

When to use it

  • Questions about recent events, prices, products or regulations
  • Explaining to colleagues why a chatbot gave an outdated answer
  • Choosing between models for a task that depends on recent knowledge

When not to use it

  • Stable knowledge such as grammar, math or long-settled history, where the cutoff rarely matters
  • Assuming a newer cutoff means the model's facts are correct