Reusable Evidence Boundary
KV Cache Prefix Reuse vs. Full Recomputation
Mode: Prefix Reuse
Pause
Reset
ZONE 1: EVIDENCE INPUT
EVIDENCE DOCUMENT
LENGTH: 4,000 TOKENS (STATIC)
ENCODER FORWARD PASS
COMPUTE KEY/VALUE MATRICES
ZONE 2: KV CACHE BOUNDARY
TOTAL TOKENS PROCESSED
4,050 TOKENS
SAVINGS: 85% RECOMPUTE AVOIDED
LOCKED KV CHECKPOINT
PREFIX KEYS/VALUES FROZEN
INDEX: [0 .. 3999]
PREFIX BOUNDARY
ZONE 3: SUFFIX EVALUATION
SUFFIX QUESTIONS (Q1..Q3)
DYNAMIC QUERY: ~50 TOKENS EACH
ATTENTION ENGINE
FETCHES CACHED PREFIX KV
GENERATED RESPONSE
STREAMING OUTPUT TOKENS
CACHE KV
REUSE KV
QUERY Q(i)
Node Title
Description text goes here.