Twenty real context chunks, embedded with all-MiniLM-L6-v2. Toggle chunks and watch the eigenvalue spectrum, the entropy, and the effective rank respond. Every number is computed live in your browser from the actual similarity matrix.
| Method | Kept | Coverage | Eff. rank | Tokens |
|---|
Each cell is the cosine between two chunks. Zero is the neutral midpoint — unrelated chunks sit near it, so the bright blocks are the redundancy. Hover any cell for the pair.
Eigenvalues are the squared singular values of E, sorted largest first, and they always sum to n because the diagonal of S is exactly 1. Dividing them by n gives a probability distribution for free — that is what Shannon entropy consumes. A tall first bar with a long flat tail means the context says one thing repeatedly.
Effective rank is exp(H) where
H = −Σ pᵢ log pᵢ and p = λ/n.
It answers “how many genuinely distinct things are in this context?”
— a continuous count, not an integer. Twenty identical chunks score 1.0;
twenty unrelated chunks score 20.0. Eigenvalues are computed with a cyclic
Jacobi rotation solver; the similarity matrix was precomputed once from the
real encoder, and every subset you select is a principal submatrix of it.