In one sentence
A knowledge cutoff is the date after which a language model has seen no training data, so anything newer must come from live retrieval or it will be missing.
What it means
Every model has one, and vendors publish it for some models but not always. A model with a cutoff last year knows nothing of your launch this year unless it can search the web. Products that search will often find it. Products or modes that do not will answer from what they had.
This explains why the same question can produce different answers in different products, or in the same product with search on and off. It also explains why a rebrand or new product can be described in outdated terms.
What to do about it
Assume newer facts reach an assistant only through retrieval. Publish them on crawlable pages and, where you can, on third-party sites the assistants already draw from. Do not expect a model's trained knowledge to update until its next release. See training data vs live retrieval.
How Pineprompt measures it
Pineprompt tracks answers daily, so a change of description after a model release or a new source becoming citable shows up as a shift in the time series.
Frequently asked
- What is Knowledge Cutoff?
- A knowledge cutoff is the date after which a language model has seen no training data, so anything newer must come from live retrieval or it will be missing.
- How does Pineprompt measure Knowledge Cutoff?
- Pineprompt tracks answers daily, so a change of description after a model release or a new source becoming citable shows up as a shift in the time series.