Anyone can bolt a chatbot onto a website.
Tokenisation
A model never sees your sentence. It sees integers from a learned vocabulary — and this one learned from nothing but the 23 chunks of content on this website.
- naive character split
- 69
- learned BPE (300 merges)
- 30
30 tokens · 69 chars
2.30× vs naive
= 2.30 chars/token
Trained on 10,444 characters it reaches 2.13 chars/token on its own corpus, from a vocabulary of 382. A production tokeniser manages roughly 4 on English prose with a vocabulary of about 100,000 — the gap is entirely the size of the corpus it learned from, not a difference in the algorithm.
Embedding space
Every chunk of this site as a point in 384 dimensions, flattened to two by principal component analysis. The two axes hold 25.1% of the variance, and the instrument says so rather than implying the picture is the whole truth.
Your text is placed at the centroid of its closest matches by term overlap — the map itself is real bge-small embeddings, projected with PCA. Your words are not embedded, and pretending otherwise would be the one dishonest thing in this instrument.
The two axes are the directions of greatest variance in the 384-dimension embedding space, found by PCA. Together they hold 25.1% of the corpus's total variance — so roughly three quarters of the structure is in dimensions this flat picture cannot show. Clusters here are real; exact distances are not.
Sampling
Temperature and top-p, over a fixed set of logits. The distribution, the nucleus, the entropy and the draw are all real — only the logits are hand-set.
- tokens kept
- 5 of 10
- mass kept
- 92.7%
- entropy
- 2.09 bits
Entropy is the number the temperature slider is really moving: at 0.1 the distribution collapses onto one token and the model becomes a lookup table; at 2.0 it flattens until lasagne is a plausible thing to build.
> Dheeraj builds — nothing sampled yet
How this was built
The embeddings are real bge-small-en-v1.5 vectors from this site's retrieval index. The projection is principal component analysis implemented from scratch — power iteration on the 23×23 Gram matrix, then deflation — and the byte-pair merge table was trained here too: 300 merges over 10,444 characters. No machine-learning library, no tokeniser package, and nothing on this page fetched at runtime.
What is not real: nothing embeds your typed query, because doing that in the browser means a model download measured in tens of megabytes. The map places your text lexically and labels the fact plainly. The reasoning behind decisions like that one lives in the decision log.