RAG copilots can't count aggregates — let the DB do it

The real fix isn’t the name—it’s a custom reducer that raises when two nodes write the same key in one run. That makes the graph fail loudly instead of silently clobbering a field until some number looks wrong. Naming conventions are just memory games.

Also add statement_timeout and read-only roles—otherwise your copilot can accidentally hurt prod.

That’s the way I’ve landed too. I used to let the model do counts and max dates, then realized a simple SQL call is faster and bulletproof. Now I keep the LLM for summarizing and explaining query results—way less hallucination risk, way more trust. Glad to see this hybrid pattern getting traction.