The executive summary gets a major update. Headline numbers: interactive chat latency dropped from 5,800ms to 500ms by fixing a provider-pinning bug in the chat service. The faster provider is now the default routing target. KB embedding model swapped from the large model to the small model — approximately 85% cost reduction with comparable retrieval quality. Qdrant integrated for approximate nearest-neighbour search. KB fully re-embedded and synced.
The AI routing matrix is now complete across all 11 segments with database settings and full controller support. Three segments were previously missing from the routing matrix, meaning form saves for those segments were silently dropped. Fixed.
Chat widget visitor identity now persisted via browser storage so returning visitors auto-rejoin their open conversation. Admin chat list defaults to open conversations instead of closed ones.
5800ms to 500ms is not a minor improvement. That is the difference between "AI feels slow" and "AI feels instant." The provider-pinning bug was a single incorrect constant that had been there for weeks.