But how? I didn't find any research paper or specifications in their developer pages.

How can an LLM be trained on realtime data? If not than how does the pplx-70b-online be online?

In my understanding, it's just doing a function call for all the queries using Bing and replying with citations. Nothing new even, Mistral and ChatGPT-4 does it. Also if vector embeddings are being used how can it be scalable?