We added reflex policy hooks around the RAG lifecycle. Policies can run before retrieval, after retrieval, and after generation.
This creates a control surface for behaviors such as warning on low coverage, refusing low-confidence paths, blocking prompt injection patterns, and enforcing token budgets.
Policy positions
The hooks are placed where they can stop different classes of failure:
- Pre-retrieval: reject unsafe or malformed inputs before they create expensive work.
- Post-retrieval: detect empty context, low coverage, or weak evidence before synthesis.
- Post-generation: check whether the answer drifted away from the selected evidence.
Design constraint
Policies must surface explicit reasons. A policy that silently rewrites behavior is another hidden prompt layer.
Why this matters
Agentic RAG needs controls that are close to the reasoning loop. Policy checks after the final answer are too late for many retrieval failures.