The Cite functionality does RAG on a .bib file to try find citations for some selected text or a given search query. It then uses the LLM as a judge, after finding the top candidate citations via RAG, so that the LLM can decide the final ranking of best options.
It uses a totally different embedding model for the RAG part, and a database of cached embeddings that are stored in IndexedDB in your browser, organized by embedding model name in case you have multiple.
Ultimately I am a bit skeptical of it, though, since the benefits of having a good handle on the literature are much more than just being able to recall papers to cite.