Troubleshooting
Edge Troubleshooting
Diagnose access, model, document, and update problems
Table of Contents
About this guide
Begin with what changed and which part of the workflow fails. Check network access, core service health, and model readiness before changing configuration. A working home page does not establish that the model or document pipeline is ready.
Access and accounts
| Symptom | Check | Next step |
|---|---|---|
| Kit page unavailable just after power-on | First startup can take up to about 15 minutes. Check required units and network equipment have power. | Follow kit startup checks; verify network cables before using the supplied restart procedure. |
| Workspace hostname cannot be found | Workstation network, DNS, and the address in the handover record. | Ask the network administrator to verify resolution to this deployment. |
| Certificate warning | Correct hostname, device time, certificate expiry, and issuer trust. | Have the administrator correct the certificate or distribute the approved trust chain. |
| Workspace opens but sign-in fails | Correct Edge instance, provisioned account, password capitalization, and identity-provider reachability. Use the current password if the credential-card password has been changed. | Contact the workspace administrator; cloud and Edge accounts can be separate. |
| Manager returns access denied | Manager account and required role. | Use the appropriate named account or have an admin adjust access. |
| Manager is unreachable after an update | An update may be restarting the Manager. | Allow it to return, then refresh. If it stays unavailable, use the supplied console procedure. |
| Wi-Fi drops after a hotspot change | SSID and password may have changed. | Reconnect with the new settings or use the prepared wired connection. |
Models and performance
| Symptom | Check | Next step |
|---|---|---|
| No local models in the picker | Inference mode, model readiness, and user access. | Load an approved model in Model Management, then refresh the workspace. |
| Model is in the catalog but cannot answer | Files may be absent or the model may still be loading. | Check its state and serving workload before setting it as default. |
| Manager asks to replace another model | The requested model does not fit the available node budget. | Coordinate unloading another model or choose a smaller model. |
| Model loading fails | Node health, memory budget, complete files, and engine compatibility. | Read Live events or workload logs; retain the error for support. |
| Responses are slow | Model size, prompt length, concurrent requests, and resource telemetry. | Try a short request with a smaller approved model and compare results. |
| One node is unavailable | Power, regular network connection, and interconnect. | Restore connectivity and wait for readiness before loading more models. |
| Vision or tool requests fail | The selected model and serving engine may not support the capability. | Use a compatible approved model and test the capability separately. |
Do not force a placement by reducing its memory reservation. On systems with unified memory, memory exhaustion can affect the whole node.
Documents and voice
| Symptom | Check | Next step |
|---|---|---|
| File uploads but answers do not use it | Processing completion, selected context, dataset permissions, and retrieval-service health. | Test with a short document and a question answered explicitly in its text. |
| Dataset ingestion fails | Storage capacity, parser, embedding, and vector-database status. | Have an operator inspect the affected services and capture the error. |
| Microphone does not work | Browser permission and availability of the local speech service. | Grant the expected permission and test speech separately from chat. |
| Voice playback fails | Browser audio settings and the local text-to-speech service. | Test a short response and have the operator check service health. |
Updates and disconnected operation
| Symptom | Check | Next step |
|---|---|---|
| Updates reports Offline | Whether the site is intentionally isolated. | Use approved update media; local chat can continue without update connectivity. |
| Expected offline release is not offered | Correct platform, complete media, and release staging. | Follow the media-specific staging procedure with the deployment team. |
| Installation fails partway | Progress output, storage, media/network availability, and node health. | Retain the error and resolve its cause before retrying through the approved procedure. |
| Version does not change after a release install | Installation completion and current Manager view. | Refresh, inspect the failure output, and contact support if the discrepancy remains. |
| Services recover but chat fails after an update | The required model may not have reloaded. | Load it in Model Management and test chat and retrieval. |
| Local chat works but a connector fails | The connector’s destination and authentication dependencies. | Use an approved reachable service or disable the unavailable integration. |
| Cloud model fails in Frontier mode | Provider credentials, endpoint/region, permissions, quota, and connectivity. | Review Providers and the provider-side configuration with an admin. |
See Updates and rollback for software regressions and Recovery planning for sustained outages or data loss.
Get support
Use your agreed Edge support channel or the Ask Sage Support page. Prepare:
- Platform, installed version, and affected node or service.
- Connected or air-gapped operating mode, and Local or Frontier inference mode.
- The action attempted, the exact error, and its time and time zone.
- What changed immediately beforehand.
- Model identifier and whether the model is ready, if relevant.
- A sanitized screenshot or excerpt of the relevant Manager progress, health, or event output.
- Troubleshooting steps already attempted and their results.
Share only material approved for that support channel. Remove credentials, recovery keys, provider secrets, private document contents, and other restricted data. For isolated sites, use the authorized transfer process instead of connecting the appliance to an external support service.