AI Cookbooks: Recipe for dynamic model selection with AI Gateway

Help documenting a pattern for routing LLM requests to different models based on content. A "model selection" route uses the Prompt Decorator plugin to ask a cheap LLM to classify the incoming query and return a model name. DataKit then rewrites the request body to set the target model before proxying to the real upstream. Semantic Cache sits on the final routes to reduce token cost.

A "model selection" route runs first. It uses the Prompt Decorator plugin to append a classifier instruction to the request and sends it to o3-mini. The response is a single model name. DataKit reads that response, rewrites the `model` field in the request body, and routes the modified request to the appropriate upstream. Semantic Cache sits on the model routes to avoid redundant calls.

```yaml
plugins:
- name: ai-prompt-decorator
config:
prompts:
append:
- role: user
content: |-
Based on the user's question, select the most appropriate model to answer it. The options are:
- gpt-4o: best for complex questions and nuanced answers
- o3-mini: best for straightforward questions and concise answers
Only respond with the model name, either "gpt-4o" or "o3-mini".
```

DataKit then calls the selection route, extracts the model name from the response, and patches the outgoing request before forwarding to the real LLM route.