[NeutralForest]

Anything Sam Rose is worth reading: https://ngrok.com/blog/prompt-caching

but the implementation will be up to your provider and harness, for deepseek, they expose some numbers: https://api-docs.deepseek.com/guides/kv_cache/ and Anthropic has a list of actions invalidating your cache: https://platform.claude.com/docs/en/build-with-claude/prompt...

Basically, you avoid anything dynamic: model change, tool change, etc it's also important that your system prompt or main prompt doesn't have non-static data like the date/time/place or someone's name (the person you interact with in a chatbot for example). That should be left to tool call or search.

[samwho]

Sam Rose here. Thank you <3

[NeutralForest]

The man himself, thank you for the articles =)

[samwho]

You are extremely welcome.

[insane_dreamer]

samwho? samrose.

[arjie]

I just put the varying parameters in a trailer prompt and have them change every time. It doesn’t matter because the cache is prefix keyed. You lose caching for the last 20 tokens or so but that’s not a big deal. Moving it to a tool call makes it too slow (needs full roundtrip).

If you’re constructing the prompt you don’t have to jam everything together you can arrange it appropriately.

[NeutralForest]

Yes indeed! Mostly don't put changing data in the beginning or prepend.