Make repetition do less work.

ILLUSTRATIVE FIXTURE
REQUEST → CACHE → MODEL
EXACT KEYS / LRU EVICTION

Canvas unavailable. The exact cache computation and export controls remain usable.

PREDICT / RUN / EXPLAIN

How many requests skip generation?

Run the same six exact prompt keys through an LRU response cache. A hit returns a stored response; a miss invokes the model and stores the new result.

ABACAB

CACHE CAPACITY

PREDICT CACHE HITS

HITS–

MODEL RUNS–

HIT RATE–

Prediction required.

LRU keeps the most recently used keys and evicts from the old end when full.

Illustrative exact-response cache only. No live API prices, latency, GPU measurements, semantic similarity, provider behavior, or universal savings are claimed.

Trace the saved cache fixtures and their explicit limits

Your browser does not support video.

Read the explanation

The conveyor uses six fixed exact keys A, B, A, C, A, B. With two LRU slots, A hits at its third and fifth positions but B must regenerate at the end after eviction. With three slots, the final B remains present and becomes an extra hit. Bars show two versus three skipped illustrative generations at one hundred seventy pixels each. This is a local fixture calculation; no live model or provider call runs. Every miss follows its illustrative model path and stores the exact key. Across six requests, two-slot capacity yields four misses while three-slot capacity yields three. Both bars depict assigned model-run counts at one hundred twenty pixels per run. These are not measured inference costs or latency; the page explicitly limits itself to exact response caching. LRU removes the least recently used end when a missing key arrives at full capacity. In the written sequence, two slots evict B before C and then C before the final B, for two evictions. Three slots retain all A, B and C and evict nothing. Bars show two versus zero evictions at two hundred pixels each. A hit moves that key to the most recent end. Inspect exported before and after cache arrays to follow the exact policy.