TITLE: Real chat traffic through a semantic cache 962 recorded WildChat turns from 176 users in three tenants Every request's decision on the record, with its reason Retire a model version, and its answers leave in one wave Recorded WildChat traffic, replayed through Grafferene. Each dot is a request. Answered from the cache, or sent on to the model. Three tenants of real users. Answers from the model are cached, each with the model version that wrote it. Click any request for its trace: how it was decided, and why. This one: no cached question close enough, so it went on to the model, and its answer was cached. Ask your own question. It is decided in your browser by an 8-bit copy of the recording's embedding model. Nothing you type is sent anywhere. Served: the same question in other words, similarity 0.94 against 0.92. Cached for: What are the most fun things to do in Southampton? Nothing cached is close (similarity 0.23): it would go to the model. This demo calls no model. Tenant A retires a model version. Every answer built from it is invalidated in one wave. 433 removed; 2 under a retention hold kept out of service. Other tenants' answers built from the same version are reached and left alone. The change is not theirs. A handbook page changes for Tenant B: its answers stay, in Probation, marked as possibly out of date. 252 answers. END: demo.grafferene.ca Recorded traffic, replayed in your browser. Measure your own: grafferene.ca