Caching

There are two cache layers. Both are keyed by URI with TTL-bounded freshness, and only validated prompts are ever cached.

L1: client

client = PromptClient(host="localhost:8443", cache_ttl=30)

Repeat gets within the TTL never touch the network. Pair it with subscribe() for immediate updates, with the TTL as the fallback.

L2: server

This layer is in-process by default (-cache-ttl 30s; 0 turns it off). For multiple nodes, share one Redis:

priompt serve -redis-url redis://localhost:6379/0 -cache-ttl 60s

Keys look like priompt:cache:<uri>. Publishing or rolling back deletes the key (write-through invalidation). If that invalidation fails, the publish returns UNAVAILABLE rather than letting nodes serve stale content.

Every cache read is best-effort: a Redis hiccup becomes a cache miss, never a serving error. When encryption at rest is on, Redis values are sealed with the same key, because Redis snapshots to disk too.