Repository navigation
Troubleshooting
An empty visible answer accompanied by reasoning and a length finish reason means generation stopped at a length limit while still reasoning. This is distinct from a network failure or a response timeout.
- Check Maximum output tokens in the panel and on the secure data-source instance. The smaller value is sent to the provider; the provider may impose a smaller limit again.
- Increase the panel budget within the administrator ceiling and model's supported limits. For example, 8,192 is a bounded starting point for testing a reasoning model, not a guarantee that it will finish.
- Use Requested answer max tokens (soft) for a concise visible answer, understanding that this is a prompt instruction rather than a separate answer budget.
- If necessary, use a provider-supported non-reasoning configuration or a different model. The gateway does not currently expose a reasoning-effort control.
The panel's provider-default output option removes only its own cap. It does not bypass the administrator ceiling. A model advertised with a 256K context window does not necessarily support a 256K output. Do not repeatedly retry with an unlimited budget or treat partial reasoning as a completed operational assessment.
Confirm the full browser URL and installed plugin version on the exact Grafana instance being tested. Replace both complete plugin packages and restart Grafana after metadata changes; replacing only module.js under an unchanged version can leave an already-loaded browser module in use. Reload the whole browser page, not just Grafana's dashboard query-refresh button. Brave's reload-ignoring-cache shortcuts are Ctrl+Shift+R or Shift+F5.
If a hard reload does not help, compare a fresh private window with the normal window before assuming a server defect or deleting browser data. A long paragraph appearing as a generated image name, rather than the panel heading, indicates older capture code or a different installation; record the full URL and both plugin versions. The 1.1.0 panel bounds generated capture labels and excludes the text-panel body. Do not change secrets or provider settings to repair an image-name validation error.
Only four requests may run concurrently per data-source instance. Additional requests fail immediately with HTTP 429; they are not queued with their image payloads. Wait for an active request to finish and retry. The separate 30-requests-per-minute limit still applies, including rejected attempts.
Confirm Allow insecure HTTP is checked on the same data-source instance and select Save & test. If the checkbox is missing or the old error remains, rebuild both plugin halves and restart Grafana:
npm run build
mage -v build:linux
docker compose restart grafanaA backend executable from an older build can continue enforcing the previous locality-only rule even when the frontend displays the newer setting.
The requested model must match allowedModels exactly. If the list is empty, only defaultModel is permitted. A nonempty list must include the default. Add the model administratively or use the configured default; secure model discovery shows only permitted IDs.
The administrator must enable Enable image analysis and approve the selected model in Allowed vision models, which must also satisfy the ordinary model allow-list. A text model is not automatically a vision model. If the provider rejects a validated image request, confirm actual vision support and reduce image count/dimensions or prompt size. PNG and JPEG are supported; remote image URLs, SVG, WebP, GIF, malformed base64, truncated images, and excessive sizes are rejected before sending.
Data-only requests do not require vision. Images-only requests still require a prompt and the secure AI connection; turn off Include query data with images in the panel if query rows should be omitted. Combined requests send only the Intelligence panel's query results, not every pictured panel's raw data.
The preview adapter uses an approved published agent's Direct Line channel secret; a normal Microsoft/enterprise SSO login is not that credential. It does not implement delegated end-user authentication, OAuth-card interaction, or SSO token exchange. Do not disable agent authentication to make a test pass. Ask the tenant administrator for an approved integration and authentication design.
The agent must return a correlated text reply and a supported completion signal (acceptingInput, expectingInput, or endOfConversation). Progress-only or non-text replies are not treated as final assessments. For images, the published agent must permit file uploads. A successful Save & test checks token exchange only, not the agent's inference, file understanding, or tenant-specific authorization.
Use the fixed resource base ending /api/plugins/grafana-llm-app/resources/llm/v1, preserve your Grafana subpath, and select the base or large alias. Store a Grafana service-account token with LLM app access in Bearer token. The configured app requires Grafana 13+; its two-minute upstream timeout can be shorter than the data-source timeout. Interactive enterprise SSO or browser login does not provision a backend service-account token.
The panel, data source, proxy, provider, and Grafana LLM app can each have their own deadlines. The first reached deadline ends the request. Reduce input/image size or use an appropriate model; review configured timeouts with the administrator. Increasing a panel timeout cannot extend an upstream provider deadline.
Remember that the backend runs from the Grafana server/container. Test DNS and connectivity from that network namespace. Redirects are intentionally rejected; configure the final provider URL directly.
Saved secrets are write-only. Select the field's reset control, enter the replacement, and save. Grafana will display only configured state afterward.