Based on the retrieved codebase memory, here's how query routing works:

Routing Logic:

1. Explicit Override: The `_should_use_cloud` function first checks for an explicit
   `use_cloud` parameter that can force cloud mode (DeepSeek) regardless of other
   conditions #main.py:986-1005.

2. Keyword Detection: If the user query contains keywords like "architecture",
   "explain", "how does", or "overview", it routes to DeepSeek's cloud mode
   #config.py:43-44.

3. Word Count Threshold: Queries exceeding 40 words automatically route to
   cloud/DeepSeek via `_should_use_cloud`'s word count check #main.py:992-994,
   with this threshold defined in config as `CLOUD_ESCALATION_WORD_COUNT = 40`
   #config.py:37.

4. Default Mode: Falls back to whatever is set as `DEFAULT_MEMORY_MODE`, which
   defaults to "hybrid" unless explicitly changed via the `MEMORY_MODE` environment
   variable #config.py:31-32, with this function returning based on whether that
   mode equals "cloud" #main.py:996-998.

Important Note: The retrieved context does not contain explicit information about
Ollama's role in the routing decision or how local model queries are constructed
when cloud mode is NOT selected.