Four things you can poke at, straight from the workshop. Nothing to install, no account, works offline. Everything you see here came from real models.
All clinical text is invented for teaching. Nothing here is medical advice.
Before a model does anything, your text is chopped into tokens — pieces from a fixed list of about 50,000. Common words get one each. Rare ones get shattered.
To compare text, a model turns each phrase into a point in space and measures the distance. 1.00 means identical. Pick a pair and see whether it agrees with you.
The model does not produce a word. It produces a score for every word it knows. A setting called temperature decides how adventurous the final pick is.
Left: it always says the same thing — repeatable, and dull. Right: unlikely words become live options — varied, and more often wrong. Nothing about the model changed. Only the dial.
Before blaming a model for poor search results, check something simpler: how much of each document did it actually read?
The popular general-purpose model accepts 256 tokens. Our notes average over 500. It read the first half of each one and discarded the rest — with no error and no warning. The half it threw away holds the hospital course and the discharge medications.