I have the same concern as other commenters (using a separate embedding model, utility of cosine similarity, etc.)

BUT this could “seed” a really neat loading graphic for reasoning models, beyond seeing the thinking steps.