Tool Routing¶
When a chatbot has tools, a model has to decide for each message whether a tool applies and with which arguments. This router is never the chat model the user picked; it is chosen from the chatbot's providers in a fixed order.
The tiers¶
- The chatbot's OpenAI Compatible provider — used when the user's selected model belongs to that provider; routing then runs against the same endpoint and model.
- The chatbot's OpenAI key — when OpenAI is enabled on the LLMs page, routing uses that key with OpenAI's
gpt-4.1. - The hosted default router — an NCSA-hosted model provided by the deployment, used when the chatbot has neither of the above.
If none of the three is available the chatbot is Offline for tools: conversations still work, but no tool is ever invoked.
What the Tools page shows¶
The badge at the top of the Tools page tells you which tier is in effect:
| Badge | Meaning |
|---|---|
| Custom router | Tier 1 or 2 — one of the chatbot's own providers is routing. |
| Default router | Tier 3 — the deployment's hosted router. |
| Offline | No router available; add an OpenAI or OpenAI Compatible provider on the LLMs page. |
For operators¶
The default tier exists only when the frontend has NCSA_HOSTED_VLM_BASE_URL set; TOOL_ROUTER_MODEL_ID selects the model, falling back to the platform's current default NCSA model. On the hosted site the default router is always available; on a self-hosted stack see the Configuration reference.