We added a fingerprint cache path to reduce repeated work when inputs are equivalent. This is not an accuracy feature by itself; it is a latency control.
The constraint is correctness. Cache keys must preserve enough context to avoid reusing an answer or retrieval result where the surrounding conversation changed the meaning of the request.
We tightened the Agentic RAG path around avoidable work: repeated decomposition, repeated embeddings, and unnecessary retrieval fan-out.
Ingestion guardrails were added to make document handling and chunk quality more explicit before retrieval uses the content.
A feedback store was added so retrieval and answer behavior can be improved from recorded outcomes instead of anecdotes.