What is Routing?
Model routing, told as a story about a guide parrot at the door that looks at each question and decides whether the small, medium or big parrot should answer.
Every question goes to the biggest parrot — even a hello. The bill explodes.The big parrot spends many beans per answer. If a hello, a summary and a contract review all cost the same, the bean bill piles up fast.
Each question needs a different size of parrot. One bird for everything means waste or shortfall.The big parrot knows more, is slower and pricier; the small one is fast and cheap but short on hard work. Give a hard one to the small bird and a plausible wrong answer goes out.
Routing is a guide parrot at the door looking at each question and picking the small, medium or big parrot.The guide parrot reads the question and tags it easy, medium or hard. Sent by tag, easy ones get answered cheaply and hard ones properly. The guide itself is a tiny bird, or just a few lines of note.
Send each question to the right size and the answers stay the same while the beans drop a lot.The small parrot is also quicker — a shorter answer time. In a multi-step errand list, each step can pick a different parrot.
The guide gets it wrong too. Mistake hard for easy and quality leaks. So fallback and tests go with it.When the small bird says not sure, pass it to the big one (fallback). And the examiner measures how far the small bird can go. If leaked quality outweighs the saved beans, you lose.
Routing = a guide parrot at the door decides whether the small, medium or big parrot answers each question. Easy ones cheaply, hard ones properly. The guide errs too, so fallback (not sure → big bird) and tests go with it.
Classifying each request by difficulty or type and sending it to the right model (small, medium, large) to cut cost and latency. The classifier is a small model, rules, or a prompt; when the small model is unsure it escalates to a larger one (fallback, cascade). An evaluation set decides how far the small model can go, and you keep checking that quality lost on misrouted requests doesn't outweigh the savings.
When grown-ups say it
- Model routing
- The guide parrot. Looks at the question and picks the small, medium or big parrot.
- Classifier
- Tagging questions easy or hard. A tiny parrot or a few rules. It has to be fast and cheap.
- Fallback
- Passing it up when unsure. When the small bird says it doesn't know, the big bird answers again. It catches the guide's mistakes.
- Cascade
- Small bird first, then the next. Escalate small → medium → big. Most questions end at the small bird.
- Cost optimization
- Saving beans. Not using the big bird for easy questions is the biggest saving. → the bean bill
- Latency
- Time until the answer arrives. The small parrot is much faster. → what answering costs and takes
- Evaluation
- Measuring how far the small bird goes. Route without tests and you won't see quality leaking. → the examiner's scorecard
- Workflow
- A different parrot per step. Each step of the errand list picks its own bird. → the errand list