Thanks for useful unpolite feedback, and yes, opensource community can provide fast and free support at some extent

Not ridiculous as long as there are still who need to learn. right? That said, I didn't see that coming

I could've done better with research, but this post has been collecting dust in the drafts, so I decided to try my first (and last) time to finish the work I started a few months ago.

Users shouldn't need to remind the LLM not to fabricate data when it doesn't have access to a specific dataset.

Thanks @GICodeWarrior for taking time commenting on the article. Shamefully, I can already imagine a scenario on how the attack could be carried out. Fortunately, the vulnerability can be corrected by introducing escapejs template filter. Big thanks to @gynvael.

Encoding for each scenario can be quite complex unfortunately. Django does have some template filters to help.

I recommend following the documentation carefully, and using a JSON API or other similarly standard mechanism if the documented options are insufficient.

I had an opportunity to work for Outlier as an AI trainer/reviewer, the point was to beat the AI models. The role doesn't relate to SWE duties, but it was a good money. Like someone swimming in the ocean when he finds a small tiny isle and takes a break for next miles.

This trend that wants clickbait to not be considered clickbait.

I love my Garmin Fenix 6 pro.

Childish AI, I gave him two photos, both was totally wrong and hundred thousands miles from actual locations.

Don't recommend.