We started the Agentic RAG track with a constraint: retrieval, evidence selection, reasoning, and answer synthesis must stay separable and inspectable.
The first working direction focused on keeping the system explicit. The runtime needed to expose why a retrieval path was chosen, what evidence was used, and where uncertainty remained. That rule became the baseline for the work that followed.
Initial engineering boundary
We framed the first pass around four separable stages:
- Retrieval: find candidate context without pretending every result is evidence.
- Evidence selection: decide what context is strong enough to use.
- Reasoning: operate over selected evidence and conversation state.
- Answer synthesis: produce a response that can be traced back to the selected evidence.
Engineering note
This milestone did not produce an accuracy or latency claim. It established the runtime boundary that later implementation work had to respect.
What we deliberately avoided
- No one-shot top-k framing.
- No hidden fallback path that silently changes behavior.
- No public claim that cannot be tied back to code or tests.
- No exposure of private implementation details.