RAG at the proxy, zero lines in the app

The support bot answers from the model's general knowledge, which means it answers about your product the way a stranger would.

RAG is the standard fix, and the standard fix means touching the app: retrieval code, embedding calls, prompt assembly. Instead, you do it at the traffic layer. The app's AI requests route through a hoody-exec script; the script looks at the question, pulls the relevant sections from your docs, and prepends them to the prompt before forwarding to the gateway.

Retrieval became infrastructure: swappable, testable, versioned as a file: while the application shipped exactly zero new lines.

The app asks a plain question; the model hears an informed one.