Sokudo
How It Works
From tokens to a reusable KV cache
01 · Tokenize
Prompt becomes token IDs
Text, whitespace, and ordering determine the exact sequence.
→
02 · Prefill
Model processes the prefix
Attention layers compute keys and values for every input token.
→
03 · Store
Prefix state is cached
The reusable state is associated with the exact prefix and cache policy.
→
04 · Reuse
Matching request skips prefill
Only uncached suffix tokens need fresh computation.
[system + tools + stable docs + history]+[new user turn]