Purrx

Cache decks and re-theme without AI

Never pay twice for the same deck: an LRU store keyed by a stable hash of the request, and theme changes that re-run only the compiler on saved SlideML.

12 min+35 XPHands-on

People regenerate the same deck constantly: a page refresh, a double-click, a colleague with the same brief. And they change the theme after seeing the result. Neither should cost another 30 seconds and another few thousand tokens.

Request

{ topic, slideCount, theme }

requestKey()

same fields → same key

store.get(key)

hit: return it

miss: generate

then store.set(key, deck)

saved.xml

the SlideML

compileDeck(xml, 'dark')

no model call

A repeated request is answered from the store in milliseconds. A theme change re-runs only the compiler on the saved SlideML, so it costs no tokens at all.

A stable key

{ topic, slideCount } and { slideCount, topic } are the same request, but JSON.stringify writes them differently. stableStringify sorts object keys first, then requestKey turns that text into a short . Same request, same key, whatever order the fields arrived in.

An LRU store

Memory isn't infinite, so the store keeps a maximum number of decks. When it's full, it drops the one used longest ago: an . A JavaScript Map makes this short, because it remembers insertion order:

  • get: delete the entry and set it again, which moves it to the end: "most recently used".
  • set: put it at the end. While the map is too big, delete the first key: the least recently used.

Re-theming for free

Because the pipeline returns the SlideML, the store keeps it. rethemeDeck(saved, "dark") just calls compileDeck(saved.xml, "dark"): milliseconds, no tokens, and exactly the same words.

theme: light

theme: dark

The same SlideML compiled with two themes. The slide only says $primary and $surface; the theme decides what those are.
Under the hood — Isn't caching AI output wrong, if each run could be different?

For a chat reply, sometimes. For a deck, the user asked for a deck, got one, and asked again with the same inputs, usually by accident. Returning what they already have is what they expect. If they want a new version, give the UI a "regenerate" button that skips the cache on purpose.

Key takeaways

  • The same request should never cost tokens twice: key saved decks by a hash of the request.
  • Save the SlideML, not just the .pptx, so a theme change only re-runs the compiler.
  • An LRU cache keeps recent entries and drops the least recently used when it's full.

Sign in to run the exercise

Reading is free. Writing code here needs an account so we have somewhere to keep your Gemini key and the +35 XP you are about to earn.

Continue with Google