AI settings
Configure models and the failover chain and the knowledge-base embedding model; see model availability, generation quality and token usage.
AI settings is in the Admin group at the bottom of the sidebar and only admins can see it.
Model overview
Model availability, calls in the last 24 hours (counted from the audit log, so auditing must be on), p95 latency, and whether any provider keeps failing.
Model routing
You can configure several providers; in list order they form a failover chain. Add a provider asks for:
| Field | Meaning |
|---|---|
| Type | openai (including any OpenAI-compatible gateway such as Volcengine Ark, vLLM or Ollama), or Azure OpenAI (needs the API version and the deployment name) |
| Base URL, model | Endpoint address and model name |
| Key | Never shown after saving; leave blank to keep the current key |
| Timeout | Per-call timeout in seconds |
| Reasoning effort | Reasoning models think before answering. Auto (recommended), off, low or high |
Click Save and reload to apply.
Multi-provider failover is an Enterprise feature: Enterprise moves to the next provider when one fails; other editions only use the first enabled provider and do not switch.
Knowledge-base embedding model
Knowledge base search needs an embedding model. Enter the model ID; leave the address and key blank to reuse the chat model's endpoint and key. Click Test connection; once it passes, the vector dimension is filled in and saving turns the knowledge base on.
To switch to an embedding model with a different dimension, rebuild the knowledge base first.
Generation quality
- Failed cases: questions a user rated down or quickly rephrased, used to extend the evaluation set and improve the prompts.
- Corrections: statements users corrected. Those the model judges to be lasting facts go into the knowledge base; the rest are discarded, and all of them are kept here.
Cost
Calls and input and output tokens over the last 30 days, per provider and model. Amounts are not calculated because prices differ between vendors; your bill is the reference.