Agent tooling
This extension adds optional safety checks to a Pi search tool, reordering results and hiding suspected harmful content.
Only you can see your notes.
Screenshot unavailable. Open the experiment ↗
The automated safety filter is not a guaranteed shield against harmful content. If classification fails or times out, the tool shows unreviewed search results with a warning. When enabled, the configured decision/classification provider receives the query and candidate titles, URLs and excerpts before suspected unsafe material is hidden; this is separate from the search provider.
Use the Pi search extension to gather results, then optionally enable judging to reorder excerpts by whether they answer your question. Ordinary search does not call Jev; judging is off by default and requires choosing an available classifier provider and model in Pi.
Enable judging only when you are comfortable sending the query and result excerpts to that provider. It estimates relevance and suspected hazards, not truth or guaranteed safety. Failed checks leave the original results visible with warnings; inspect the sources rather than treating the filtered list as cleared evidence.