# Willison's gerbil

> I indexed the docs for Simon Willison's library as an example for this RAG pipeline.

I indexed the docs for Simon Willison's `llm` library as an example for this RAG pipeline. Then it suddenly became a gerbil.

![A form interface with a query field containing "What is the meaning of life?" and a preview output showing a witty, snack-themed response to the question.](/images/_1798578490759078263-GPXVmPobgAAQ9Ii.png)

Well, turns out that one time Simon put "Pretend to be a witty gerbil" as an example prompt in the Changelog. Apparently this was similar enough to my sample question that it got included in the retrieved documents. Truly a lesson on prompt injection.

![This image shows a user interface for continuing a chat with OpenAI models, demonstrating how to use the -c/--continue option to resume a previous conversation where the AI pretends to be a witty gerbil.](/images/_1798579358917120091-GPXV1XGbcAA7mo3.jpg)

Now I've mentioned this on the internet, giving it more weight. Other people might use the `llm` docs as a sample dataset, just to get this fun effect.

We have birthed something: a new internet cryptid. [Willison's Gerbil](https://simonwillison.net/2024/Jun/6/accidental-prompt-injection/)

[View original](https://twitter.com/i/web/status/1798578490759078263)