Working with embeddings - particularly with the goal of aiding thinking - a lot of the challenge is where do you draw the line for chunks. Paragraph-by-paragraph? Essay-length? Book-length?

Makes me appreciate books as containers for thoughts. Part of their purpose that you can use them as reference-able point for 'this set of ideas'. Maybe you'd like to decompose them down into their component ideas - but that introduces judgement on the model's part - embedding the whole book (or a generated summary of the book) is more stable.

With stable points in space you can rearrange - maybe interpolate. Although there again the clean picture of it to me is in 2D space, but embedding is 768 dimensions or more. Feels like you want to flatten to do some investigation then reinflate, then flatten to investigate another angle... But you lose some stable ground then probably.