You.com generally powers the AI for our site using proprietary commercial Large Language Models (LLMs). These produce high-quality text and can reliably power sophisticated features. They also have advanced moderation features permanently built in to attempt to ensure their output is appropriate and safe for work (SFW).
However, for users who have set Safe search to Off, these moderation features can sometimes be overzealous, leading to You.com returning search results that it refuses to summarize, or similar loss of functionality.
The “Safe search: Off (uncensored chat)” option instead uses a less-censored open-source LLM. Currently we are using the new Mistral 7B model, specifically the Zephyr 7B alpha model instruct fine-tune of Mistral. You.com will likely update this choice frequently as even better open-source models come out. This LLM is not absolutely compliant: it emulates what could be described as “community standards” across the web, so there are still some queries that it may refuse, but these are significantly fewer than for commercial models. Using an open-source model also gives us the option to apply our own moderation to modify its default behavior if we find this is necessary.
Please be aware that if you ask this LLM for Not Safe for Work (NSFW) content, it will frequently provide it — much like any other search engine with Safe search Off.
While this open-source model is significantly less censored than You.com’s regular commercial AI, it is also a significantly less powerful model, so it can produce lower quality text and make more mistakes. Therefore, we don’t recommend using this Safe search option as your permanent default setting. Instead, we recommend you switch to it only when you encounter or anticipate refusals from our regular AI.
For this reason, we’ve added a notification to the end of its generations to remind you that you are currently using this setting.
