Here, I fed it notes from a horticulture class. It broke the notes apart into small chunks (roughly a paragraph), then mapped them into a 3072 dimensional space (where coordinates correspond to semantic meaning) using the gemini_embeddings_2 API. These embeddings are passed to a UMAP algorithm that reduces the dimensionality enough for a clustering algorithm, HDBSCAN, to properly group them. A Gemini Flash Lite model then reviews the content of the cluster and names them.

These embedding models are usually used in retrieval augmented generation (RAG) to allow LLMs to know where to search for context in a document uploaded, based on how closely the prompts embedded location is to what you've uploaded to it.