Companion code: MukundaKatta/ragvitals-gemma-demo . The Agentic-Postgres entry point is demo/pgai_ollama_run.py plus the agent_drift_view.sql script in this post. Most RAG agents in production are flying blind. They retrieve, they generate, they hand the answer to the user, and they have no idea whether what they just did was good. The eval lives in a notebook the data scientist runs on Mondays. What if the agent could ask the database directly: "is my retrieval quality drifting this hour"? Not as a stretched-tool-call to some external observability service, but as a plain SELECT over the same Postgres that holds the corpus? This is the experimental angle: turn ragvitals into a set of SQL functions on top of Tiger's Agentic Postgres, expose them as MCP tools, and let the agent self-audit between actions.…