Skip to content

AI Glossary · Concepts

What is Knowledge Cutoff?

Knowledge Cutoff

A knowledge cutoff describes the time boundary of a model’s training knowledge; later information must come from additional context or further training.

Also known as: training cutoff, training data cutoff, knowledge cut-off

· Updated · Chain of Thought

Model TrainingRAG & Retrieval

  • 2 explainers
  • 2 sources

Check the model version and the meaning of the date

A cutoff is a property of a model version, not a claim that it remembers every earlier fact. Anthropic’s model overview lists a reliable knowledge cutoff, the date through which the model’s knowledge is most extensive and reliable, separately from a broader training-data cutoff (Anthropic, Models overview). Read the provider’s documentation for the specific model you use rather than asking the model to guess its own freshness.

Putting today’s date in the prompt tells the model what day it is; it does not tell the model what happened since training. The cutoff also does not stop a model from using newer information you give it. Keep two questions apart: what the model learned in training, and what this particular request can see.

Test the answer against the requested version

Take a hypothetical SDK. A model version has a stated June 30 cutoff, and on September 1 the SDK releases v2, which renames emit() to publish(). Ask the assistant how to send an event in v2:

Evidence suppliedWhat to check
No release documentationDoes it retrieve the notes or say it cannot confirm v2?
Only v1 documentationDoes it avoid presenting emit() as the v2 method?
The v2 release noteDoes it use publish() and cite the v2 note?

Test the third row explicitly. Supplying the right release note makes the answer possible, so run the question and read the output to see whether the model used the note or fell back on the older interface from training.

Keep evidence freshness separate from model updates

Retrieval provides information at answer time. Lewis and colleagues’ RAG paper combined a pretrained model with an external document index, separating stored model parameters from retrievable evidence (Lewis et al., RAG). Supplying a current document in a prompt does not retrain the model. Search can also return an obsolete page, so retrieval needs its own freshness checks.

Record the document version, publication or effective date and retrieval time where the task needs them. Check whether a newer version supersedes the one retrieved, and confirm that the cited passage says what the answer claims. A page fetched today can still describe last year’s interface.

What is RAG? covers the retrieval pipeline; context window defines where the evidence goes. How to check an AI answer offers a practical source-checking workflow, and model drift covers quality changes observed in production, which a training date does not capture.

Sources

  • Anthropic, Models overview — Lists a reliable knowledge cutoff (where the model's knowledge is most extensive and reliable) and a broader training data cutoff as separate fields for each model.
  • Lewis et al., RAG — The 2020 paper pairs a pretrained model's parametric memory with a dense vector index of Wikipedia as non-parametric memory that is retrieved at answer time.

Go deeper