> Would an LLM-driven "Personal Library" require manually annotated textual interpretation of each curated item, or could it derive personal interpretations from user history and the uniqueness of curated items/sets?

I’ve personally found that tagging is less robust than LLM embeddings (mainly due to dimensionality), but human appended thoughts about a source — also embedded — serve even better as tags.

Example: “this is a quote about dinosaurs…” (Old way of doing things) Tags: dinosaurs, jurassic, history Query: “dinosaurs” > results = 1…

(New way of doing things) Embedded Quote: [0.182…] User Added Thought: “this dinosaur reminds me of a time i went to six flags with my cousins and…” Embedded User Added Thought: [0.284…]

Query: “dinosaurs” > results = 2 (indexes = sources, thoughts)

The "thoughts" index can do a second layer cosine similarity search and serve as a tag on its own to fetch similar concepts. Basically a tree search created by similarity from user input/feedback loops.