We tightened the Agentic RAG path around avoidable work. The target was not to make the system look faster in a demo; the target was to make expensive steps visible and defensible.
What changed
- Repeated decomposition paths were reviewed against cache behavior.
- Retrieval fan-out stayed tied to query complexity rather than becoming a default.
- Embedding reuse stayed explicit so unchanged inputs do not trigger unnecessary work.
- Diagnostics remain part of the loop so latency reductions do not hide retrieval behavior.
Measurement boundary
This update does not publish latency numbers. Latency claims require measured runs, environment details, and variance. This pass only documents implementation direction and control points.
Engineering note
Low latency and grounding can conflict. The system should avoid repeated work, but it cannot skip evidence selection just to save time. That boundary remains explicit.