fow-cache

semantic cache guard · live demo
click / space: next · r: reset · v: view
Net saved
$0.0000
LLM calls avoided $0.0000 − guard $0.0000
Served from cache
0
0 full LLM calls
Wrong answers blocked
0
cached answer rejected by Jev
Plain semantic cache, same traffic
Not run yet
last message only, similarity ≥ 0.90

1
User 1
fow
fow-cache guard
embedding finds candidates · Jev decides
net saved$0.00000 from cache · 0 blocked

Cache0 entries

2
User 2
Scripted replay: no live model calls. The Jev scores are illustrative, chosen to match how the guard behaves on the benchmark. The thresholds (whole conversation ≥ 0.70, rewrites ≥ 0.15) and the flow are the library's: GuardedCache.lookup → Guard.check.
Costs from fow-cache's cost model (list prices, 2026-10-03): a hit saves one Sonnet 5.5 turn with an 8k-token prompt-cached prompt, $0.0056. The guard costs $0.00023 per turn for batched facts, plus $0.00084 per candidate (two Haiku 4.5 rewrites and the Jev checks).