fix warning

Rewrite using more straightforward API: tokenize first, then compute logprobs

WIP using tokenize endpoint directly

Add a vLLM-backed show-internals interface

Try to report memory errors, and try clearing the traceback to avoid leaking the prior cache.

link to my website from prototypes

Fix next-token blank

Merge remote-tracking branch 'origin/main'

Disable revisions page

don't need api server

Maybe run this all in one

Add URL paths for subpages

Prepare to move up to Gemma3

Increment versions

Style tweaks

Work around the empty-message bug.

clean up display a bit

Add a quick and dirty show-next-token-logits page

add a border to spans

Add a quick and dirty "show internals" page

Looks like I never committed these improvements to the backend.

styling tweaks

fix another state-management bug

Remove debug print

Fix buggy Send button

extend type-assistant-response to predictive-text for type user messages too

tweak

token showing in highlights

missing deps

ignores

Show a problem with the current approach.

sync up with the backend

Make it actually a predictive-text interface

Initial stab at typing the assistant response

copy in the backend code

remove debugging print

add option to show additional alternatives, even those that aren't the most-likely-next-token

changed color, removed the title attribute

display tweaks: only show on hover

refactor: move more to styles instead of inline CSS

refactor: use a style tag instead of some inline styles

Refactor highlights display into a separate component

This is the pyproject.toml that I've been using

Allow disabling the alternative suggestions.

quick styling edits

Simplify, removing the nuance of updated_doc

Single document approach

Always start off updated doc with original doc

Use green instead of blue for suggested locations

Fix trailing spaces causing bad predictions (grr tokenization)