Add a language model connection
In AI configuration → LLM management, add a name, provider type, API base URL, model identifier, and credential. Test connection checks the current form. Include the provider's base prefix, such as /v1, rather than a complete chat endpoint path.
The current provider type uses the OpenAI-compatible protocol. Compatible gateways and self-hosted services can be configured; native Anthropic and Gemini protocols are not current provider types. Choose a model supporting tool calls to query platform data.
Enabled, default, and selected models
Multiple connections may be enabled. Ordinary questions, skills, and tasks without a selected model use the default connection; without an explicit default, the earliest enabled connection is used. A default connection must be enabled. Skills and tasks may select their own model.
LLM management is for language models. Model management under forecasting connects time-series models such as Chronos for numerical forecasts; the two connection types are managed separately.
Advanced settings
| Setting | Purpose |
|---|---|
| Request timeout | Default 90 seconds, range 5–600, for one model request. |
| Tool rounds | Default 6, range 1–12, limiting successive tool-call rounds for a question. |
| Temperature | Optional sampling setting, subject to provider support. |
| Maximum tokens | Bounds a reply; zero omits the limit and lets the service decide. |
| Context length | Context budget; older messages are trimmed first. Zero means unknown and disables trimming by this setting. |
| Proxy and TLS | Configure a proxy or per-connection certificate-verification bypass. Retain verification for normal certificate connections. |
| Custom headers | Additional gateway headers, stored encrypted without returning saved plaintext. |
| Custom JSON parameters | Additional service parameters; model, messages, tools, stream, and other platform-managed fields cannot be overridden. |
Test connection uses the current endpoint, model, and advanced settings. Editing can reuse an existing credential. Save after testing and review the default selection.
Credentials and permissions
Credentials and custom header values are encrypted and are not returned in plaintext in lists or edit forms. Connection management and chat have separate permissions.
System settings retain only the global assistant.enabled switch. Manage endpoints, credentials, and limits in LLM management. Explicit database migration moves legacy assistant parameters into connection records.
After connecting
- Test the connection and verify the model identifier and credential.
- Enable it, select the default, and ask the assistant a question requiring platform data.
- Review tool reads and the answer to verify tool-call compatibility.
- Select other enabled models for skills or tasks with different needs.