The longer version
I did not set out to learn LiteLLM, RAG, gateways, any of it. I was wasting tokens. Every question went out with the whole tool list attached, whether I asked about a virtual machine or something I could have grepped. For a while the compose file was just names I copied.
The picture that stuck: chat talks to LiteLLM, LiteLLM talks to the model. RAG is not memory. It is a search over notes I already wrote, stuffed back into the prompt so the model can quote me instead of guessing. I have to load those files myself. They do not update on their own. I found that out by getting an old answer.
The router came after I measured the waste. Keywords pick a smaller set of tools. Some two-part questions get split so the second half actually gets asked. It is string matching. I had been waiting for something cleverer, and that was the mistake.
I did not need a smarter model to spend fewer tokens. I needed to stop sending the whole workshop when the question only needed one drawer.