Configuration AI Providers

Configuration: Secure AI Provider

Panel 1.1.0 requires a configured Intelligence Gateway Secure AI data-source instance from companion plugin 1.1.0 or later. Provider URLs and credentials are configured in that data source, not in the panel. The panel stores the instance UID and non-secret analysis settings.

Secure AI data source selects the model connection. Queries → Dashboard separately selects another panel's measurements. Both are used for data-only or combined assessment; an images-only assessment still needs the secure AI instance but does not need query results.

Provider choices

Provider	Example base URL	Credential and model requirements
OpenAI	https://api.openai.com/v1	Secure API key; approved Chat Completions model accepting the adapter's generation parameters.
LM Studio	https://lmstudio.example.com/v1	Exact loaded model ID; secure credential only if the server requires one.
Open WebUI	https://openwebui.example.com or https://openwebui.example.com/api	Approved API credential; exact model ID exposed to that credential.
Grafana LLM app	https://grafana.example.com/api/plugins/grafana-llm-app/resources/llm/v1	Dedicated Grafana service-account bearer token; base or large alias configured by the app administrator.
Custom OpenAI-compatible	https://ai.example.com/v1	Compatible /models and /chat/completions endpoints; approved model and credentials.
Copilot Studio Direct Line preview	https://directline.botframework.com/v3/directline	Secure Direct Line channel secret; descriptive agent alias. Not enterprise SSO or delegated Entra authentication.

For each instance, configure the default model, approved model list, timeout, administrator output ceiling, HTTPS/network policy, and only the required secure credential. Select Save & test. A successful health check does not prove inference quality or image understanding.

OpenAI-compatible adapters use buffered Chat Completions with max_tokens and, when configured, temperature. Models requiring other parameter names, unsupported temperature settings, or the Responses API need a different adapter. A discovered model is not automatically compatible. See Image Analysis for exact vision requirements and provider limitations.

The Grafana LLM app is optional and not required for the other five choices. App 1.2.0 requires Grafana 13 and has an upstream two-minute timeout. Its dedicated service-account token is not the viewing user's SSO identity.

Copilot's agent owns model choice, tools, temperature, and output-token policy. Direct Line does not enforce the panel's temperature or hard output limit. The preview requires a compatible correlated reply/completion contract and tenant-approved secured channel. See Copilot Studio Notes; mock coverage is not enterprise-tenant validation.

Panel controls

• Secure AI data source selects the backend instance.
• Model can be typed or loaded securely. The backend returns only administrator-approved models or agent aliases.
• Maximum output tokens is a request cap from 64 to 1,048,576 where supported; the smaller panel/admin limit applies. The range does not imply model support for its maximum.
• Provider/model default output limit omits only the panel cap; it does not remove the backend ceiling or make output unlimited.
• Requested answer max tokens (soft) adds an approximate concision instruction, not a hard limit on visible text or reasoning.
• Response timeout cancels the browser wait from 10 to 600 seconds; backend, proxy, and provider deadlines can be shorter.

Reasoning models can consume their output budget before producing a visible answer. Increase both relevant hard caps within approved budgets, not only the data-source ceiling. For example, a suitable local model may be tested with 8,192 output tokens, a 500-token soft answer target, and 300-second panel/backend deadlines. These values are examples, not guarantees or a separate reasoning-effort control.

Credentials belong in secureJsonData, never dashboard options, prompts, images, or arbitrary JSON query results. See Secure Backend and Secret Storage and the data-source configuration guide.