LLM Cache
Domain: Models & Providers
Motivation
The LLM cache is provider-side KV-cache reuse of an unchanged request prefix. DeepSeek reports cache hits in its usage (cache-read / cache-write tokens), and the token meter anchors a measurement to provider usage only when the request envelope is identical. A changed system prompt, tool schema, history, or image budget can prevent reuse from the first affected token.
Concrete Example
On a systemPromptUpdate: in-history catalog entry, a changed system prompt is appended after the cached history so the earlier prefix stays reusable; usage then reports the cache-read tokens.
Analogy
A photocopy of a long document — you reuse the identical pages and only re-type what changed.