11.14 Troubleshooting
| Symptom | Check | Fix |
|---|---|---|
| Overview: Model … Not installed on this server | Release package on the server. | Reinstall/upgrade the server release; the model is part of it. |
| AI service … Not responding | Administration → Status, row AI service (ada-ml); System service logs (AI Service). | Restart the AI service from Status. Agents keep scoring; only uploads, training and tests are affected. |
| Upload fails: The AI service uses encoder X, agents use Y. Restart ada-ml… | Server partially upgraded. | Restart the AI service. |
| Upload fails: No text could be read… | Scanned image PDF, empty or protected file. | Use a document with a text layer. |
| Train button disabled | AI service, model, or no type with 5 examples. | See the banners and the hint under the button. |
Type shows not trained after training |
It had fewer than 5 examples (see the training ⓘ). | Add examples; retrain. |
| Rule never fires on any endpoint | Module off for the scope (Overview → Module); rule still in test mode; rule not published; type not in the trained layer (editor shows not in the installed model); threshold too high. | Fix and publish; check with Test a document. |
| Rule fires on Windows but not on Mac Office save | Office for Mac uses the web add-in, which does not evaluate AI. | Expected; the macOS file scan evaluates the saved file. |
| AI stopped after a product upgrade | Banner The AI model was upgraded… | Train again. |
| Endpoint status AI Classification: Preparing for a long time | Agent cannot download /ai/… over 443 (proxy, TLS
inspection). |
Allow the path; the agent retries automatically. |
| Endpoint status Unavailable | Windows service.log: lines starting AI model … not
available yet: (reason and next try), AI scoring paused for 10
minutes after repeated failures. |
Resolve the reason (disk, download, hash mismatch); restart the agent service if needed. |
| Office saves get no AI suggestion, file scan does | service.log: AI score request refused (…):
|
Fix the reason; the rule must apply to the Office module scope. |
| Already-labelled documents get no AI result | By design (section 11.7.5): no AI rule could raise their label. | Expected. |
| A burst of audit records after training | Re-evaluation of files with the new layer. | Expected; batch your trainings. |
| Too many false matches | Test-mode audit, Test a document. | Counter-examples, higher threshold, Whole document instead of or a part, combine with a pattern. |
| Too many misses | Real documents differ from examples (language, layout). | Add varied examples, especially in the second language; lower the threshold cautiously. |
The Windows service also writes a ten-minute summary (AI scoring
in the last 10 min: …) to service.log; it never
contains document text.