Martha van Berkel

Back to the Future reboot starring Robert Downey Jr and Tom Holland??

Please read the whole thing. It's just 3 points. 1. LLMs *do understand schema* when they see it in input, contrary to a popular belief that the structured data is lost due to tokenization. This is because they're trained on it as they are trained on code. 2. LLMs *don't* always see schema in input because most tools and functions supplying context to the model strip it away during pre-processing. (The next part is where SEO people get confused.) 3. The tools AI systems use to process content which is passed as context to LLMs sometimes parse the HTML page by using structured data to better understand the page and sometimes they don't. It's a software design choice and nothing to do with LLM's capability. TL;DR: LLMs understand structured data if they get it, but most of the time they get plain text / markdown, which *may* or *may not* have been rendered by leveraging schema at the pre-processing level.