Discourse AI spam detection "Scan log" is frequently truncated

Bug
1

Priority/Severity:

Low

Platform

Operating System

Windows 11

Browser

Google Chrome

Discourse

31ed2620caa780c941c6cff2a8ecaeffcbdbb14b

Description:

Discourse AI spam detection has a useful “test” feature which allows the site admin to manually trigger a scan of a specific post. The result of the scan is then shown. This is essential for refining the “custom instructions” to reduce the frequency of false positives and negatives.

:bug: The content that should be present at the end of the scan logs is frequently cut off.

Frustratingly, the scan logs always contain the entirety of the useless repetition of the custom instructions and post content, but the actually valuable part of the logs is incomplete.

Reproducible steps:

  1. Log into a forum where you have admin privileges.
  2. Open the /admin/plugins/discourse-ai/ai-spam page.
  3. Click the “Test…” button.
    The “Test spam detection” dialog will open.
  4. Enter the URL or ID of a post on the forum into the “Post URL or ID” field in the dialog.
  5. Click the “Run test” button.

:bug: The last part of the logs content where the AI explains its reasoning may be cut off.

For example:

[...] While the post *content* itself isn',t overtly promotional or nonsensical—it appears to

Additional context

I am experiencing the problem on forum.arduino.cc (Enterprise hosting plan), which is using the “CDCK Hosted Small LLM” LLM.

1 Like