Panel Setup and Configuration

Panel Setup and Configuration

Grafana Intelligence Gateway analyzes query data, images, or both. This guide covers panel 1.1.2, which requires companion 1.1.0 or later and Grafana 11.6 or later. The optional Grafana LLM app may require a newer Grafana version.

Mode	Query setup	Image setup and action
Data only	Configure this panel's queries, commonly Dashboard → Source panel.	Leave images unattached; select Analyze.
Images only	No query results or Dashboard source required.	Turn Include query data with images off; preview images, select Use images, then Analyze images.
Data and images	Connect the query results needed for precise values.	Leave Include query data with images on; preview images, select Use images, then Analyze images.

All three modes require AI provider → Secure AI data source. It connects the model. Queries → Dashboard is a separate setting that reuses another panel's query results; selecting the AI data source does not import measurements. Capturing a panel adds its rendered image, not its DataFrames.

The supplied demo instance, DEMO ONLY - Mock AI (no inference), returns a fixed receipt: no model runs and no data or image analysis occurs. Read the analysis panel's AI connection label to identify the currently selected instance. For genuine analysis, use Connections → Data sources → Add new data source to create a separate real-provider instance, configure and test it, then select it and its approved model under the panel's AI provider settings. Follow the two-phase walkthrough; the chart's TestData/Dashboard connection is not an AI provider.

1. Configure the required secure data source

Install digitalrcs-intelligencegateway-datasource 1.1.0 or later, restart Grafana, then create an Intelligence Gateway Secure AI data-source instance.

1. Select OpenAI, LM Studio, Open WebUI, Grafana LLM app, Custom/OpenAI-compatible, or Copilot Studio Direct Line preview. The Grafana app is optional; Copilot is not an enterprise SSO integration.
2. Configure the administrator-controlled base URL.
3. Set a default model and include it in Allowed models. If the list is empty, only the default model is permitted.
4. Set the backend timeout and maximum output-token ceiling.
5. Store the provider's API key, bearer token, or Copilot Direct Line secret only in its dedicated secureJsonData field. LM Studio may need no credential. A Grafana LLM app token is a dedicated Grafana service-account token, not a browser login.
6. Keep HTTPS enabled. Use Allow insecure HTTP only for a trusted local development network.
7. Select Save & test.

See AI Providers for endpoint examples, and Secure Backend and Secret Storage for provisioning and policy. OpenAI-compatible models must accept the adapter's Chat Completions parameters; model discovery alone does not guarantee compatibility.

2. Add query data when needed

For data-only or combined assessment, create or identify a panel that returns the data to analyze. Time series, tables, logs, and other query results are accepted when Grafana exposes them as DataFrames. Confirm that the source returns data for the active dashboard time range. Skip this step for images-only assessment.

3. Add the Intelligence Gateway panel

1. Select Add visualization.
2. Choose Grafana Intelligence Gateway.
3. For query-data sharing, in Queries, select the -- Dashboard -- data source. For images-only use, skip steps 3–5.
4. Under Source panel, select the panel whose results should be analyzed.
5. Choose All data unless annotations are specifically required.
6. Save the dashboard.

The Source panel title or ID (hint) is descriptive prompt metadata only. The Dashboard data source performs the actual connection.

4. Configure AI generation

Option	Purpose	Example	Guidance
Secure AI data source	Selects the backend instance by UID.	Intelligence Gateway Secure AI	Required. The panel stores only the UID.
Model	Chooses an administrator-approved model.	review-model	Use Load models securely or enter an allowed ID.
Temperature	Controls response randomness where supported.	0.2	Range 0–2. Low values can reduce variation, but do not guarantee repeatability. Not enforced by Copilot Direct Line.
Provider/model default output limit	Omits the panel request cap.	Off	The backend administrator ceiling and provider limits still apply.
Maximum output tokens	Hard panel request cap where supported.	1200	Slider range is 64–1,048,576; the backend applies the lower panel/admin cap. Model limits can be lower; Copilot owns its generation policy.
Requested answer max tokens (soft)	Adds a concision instruction.	0	Zero disables it. Models cannot guarantee exact token counts.
Response timeout	Cancels the panel wait.	300	Range is 10–600 seconds. Align it with the backend timeout.

Responses are buffered and cancellable; tokens do not stream into the panel as they are generated. Provider credentials, URLs, authentication headers, and provider error bodies never enter panel options.

For reasoning models, the output budget may include both reasoning and the visible answer. Raising only the data-source ceiling does not raise a smaller panel request: a panel set to 1200 still requests only 1,200 output tokens even when the data source permits 256000. As a bounded starting point for a suitable local model, try Maximum output tokens 8192 in both panel and backend, Requested answer max tokens (soft) 500, and Response timeout 300 seconds in both. These are examples, not guaranteed capabilities. A model's context window is not its output budget. The soft target is approximate, the default-limit option is not unlimited, and the panel does not expose a separate reasoning-effort control.

Output-limit errors and timeouts are different. The shortest applicable panel, backend, proxy, and upstream deadline wins. The Grafana LLM app 1.2.0 has a two-minute upstream timeout that these settings cannot override. Copilot Direct Line does not enforce temperature or output-token settings; configure those policies on the agent.

After analysis, the generated assessment appears with Clear analysis and Refresh assessment. Clear analysis cancels an active request and removes the answer, errors, and pending images without modifying the dashboard query or source data. A successful image assessment releases its attachments; capture or attach again before requesting another image assessment. Refreshing without new attachments is a data-only request.

5. Configure prompts and skills

Option	Purpose	Example
System / backend prompt	Defines the model's role, evidence rules, safety rules, and response format.	You are an SRE analyst. Separate observations from hypotheses.
User message template	Defines the task and injects runtime variables.	Assess {{panelTitle}} for {{timeRange}} using {{data}}.
Skills / additional context	Adds thresholds, runbooks, ownership, and domain definitions.	p95 latency critical: 1000 ms.
Constructed prompt preview	Shows assembled messages before live data is injected.	Read-only preview

Supported variables:

Variable	Runtime value
{{data}}	Bounded JSON serialization of every received DataFrame.
{{timeRange}}	Current dashboard start/end timestamps.
{{panelTitle}}	Intelligence Gateway panel title.
{{panelId}}	Numeric Grafana panel ID.
{{skills}}	Skills/additional context.
{{sourcePanel}}	Descriptive source-panel hint.

Treat dashboard field names, labels, values, and logs as untrusted evidence rather than instructions.

6. Bound data and behavior

Option	Purpose	Default
Recent rows per frame	Includes the most recent rows per frame; range 1–1,000.	50
Maximum context characters	Caps serialized DataFrame JSON; range 1,000–200,000.	24000
Source panel title or ID (hint)	Descriptive prompt label; does not establish a data connection.	Synthetic traffic
When no data arrives	Allows, warns, or blocks text-only analysis without frames. Attached images permit image analysis even with Block selected.	Warn
Analyze automatically	Runs data-only analysis after data/prompt changes; pauses while image inputs are open or attached.	Off
Auto-analysis debounce (ms)	Waits after a change before automatic analysis; range 300–30,000. Shown when automatic analysis is enabled.	1200
Show Analyze button	Keeps manual analysis available.	On
Clear analysis	Cancels an active request and removes response, errors, and pending images.	Runtime button

Start with manual analysis. Enable automatic analysis only after provider budgets, prompt behavior, and data limits are understood.

7. Configure image assessment

Image controls are included in 1.1.0 and require companion 1.1.0 or later. They are optional: keep using Analyze with no attachments for data-only analysis.

Control	Purpose	Default or example
Show image controls	Shows the runtime Images button. Any pending images remain reviewable even if this setting is turned off.	On
Include query data with images	Sends this panel's bounded query rows alongside the images. Turn off for an image-only assessment.	On
Images	Opens capture, upload, and removable previews. It does not send an assessment.	Runtime button
Use images	Closes the preview and keeps the images staged in memory.	No network submission
Remove / Remove all images	Discards one or all staged attachments.	No network submission
Analyze images	Sends the staged images and prompt through the selected secure companion.	Explicit action only
Clear analysis	Cancels a request and removes the answer, errors, and pending images.	Source measurements unchanged

The administrator must enable image analysis and approve the selected vision model in the data source. Automatic analysis is paused while images are staged or the dialog is open. Successful image assessments release the attachments; capture again for a newer view. See Image Analysis for all capture limits and privacy considerations.

8. Configure the display

Option	Purpose	Default or example
Assessment title	Heading above the generated assessment.	Intelligence assessment
Description	Optional explanatory text beneath the title. Use it to label a mock or live environment.	Empty; for example Assessment of synthetic data
Background color	Assessment background.	Transparent
Text color	Overrides the theme's text color; leave empty for theme-aware text.	Empty
Font size	Assessment font size in pixels; range 10–48.	14
Padding	Inner spacing in pixels; range 0–64.	16
Alignment	Left, Center, or Right.	Left

Keep sufficient contrast and verify both Grafana themes when choosing custom colors.

9. Verify

First analysis

1. Select Load models securely and confirm only approved models appear.
2. Select Analyze.
3. Confirm the Markdown assessment renders.
4. Compare cited fields, values, and timestamps with the source panel.
5. Select Clear analysis and confirm only generated state is removed.
6. Save the dashboard.

For images-only verification, disconnect query results, turn Include query data with images off, attach an approved synthetic image, preview it, then select Analyze images. Check known visual facts in the response. For a combined test, reconnect the query results and turn that setting on. A mock image-count receipt verifies transport, not visual understanding.

Troubleshooting

Symptom	Likely cause	Action
No secure data source appears	Companion plugin is missing or no instance exists.	Install it, restart Grafana, configure an instance, and reload the editor.
Save & test fails	Endpoint, HTTPS policy, credential, model, or network error.	Correct the data-source configuration; provider details remain server-side.
No models load	Provider /models failed or no approved model was returned.	Verify data-source health and ensure the default/allowed IDs exist upstream.
Requested model is rejected	It is outside the administrator allow-list.	Select an approved ID or ask the administrator to update policy.
No visible answer or reasoning-only response	The model may have exhausted its output cap before the final answer; empty provider output can have other causes.	Read the sanitized error, compare both caps, and use an appropriately supported model.
Request times out	Model generation exceeded a panel or backend deadline.	Reduce context, use a faster model, or align/increase both timeouts within 600 seconds.
No source data	Dashboard data source is disconnected or source panel is empty.	Re-select -- Dashboard --, choose the source panel, and inspect its active-range data.
Image controls cannot send	Companion is older than 1.1.0, vision is disabled, or model is not approved for images.	Upgrade both plugins, restart Grafana, and ask the administrator to check vision policy.
Old labels or controls after upgrade	Browser tab is still using the earlier plugin runtime.	Restart Grafana after installation, then reload the whole page; dashboard Refresh is not a code reload. Follow the cache guidance below.

See Connecting Data from Other Panels for transformations and multi-frame behavior.

For persistent stale code, see the development cache guide. Do not clear all browser credentials or enable server development mode just to bypass cache.

Grafana 12.4 dashboard capture

Use panel 1.1.2 or later for tall dashboards on Grafana 12.4.0. This fixes document scrolling so lower panels can be visited. A stalled scroll reports an error instead of producing an incomplete image. Existing size limits and partial-capture warnings still apply. Companion 1.1.0 or later remains required; updating Current View Exporter alone does not update this panel.