A routing layer above RAG, not just another model call
Abstracts
Abstracts
I revisited my piece "AI Routing: The Missing Quality Layer Above RAG and MCPs" this week, and the core argument still holds up: retrieval-augmented generation is powerful but indiscriminate, so enterprise chatbots keep answering questions with whatever the retriever hands back, hallucinated details and all. The fix I lay out isn't a bigger model, it's a lightweight "traffic-cop" model sitting above RAG and MCP calls that decides whether a request should even be answered from retrieved context in the first place.
The practical detail that matters for anyone building this: it doesn't have to be slow. The article walks through tuning that routing layer from 640ms down to sub-100ms, which is the difference between a gate that developers rip out because it adds latency and one that quietly does its job on every request.
That's the same problem I run into building agent-managed profile updates: separating what a system verifies as fact from what it's allowed to claim out loud. A routing layer for RAG and a review gate for generated updates are solving the same shape of problem, just in different domains.