This issue has been created
 
 
LLM AI Integration / cid:jira-generated-image-avatar-15fc8600-9703-4abe-a01e-4958ac1f585d LLMAI-232 Open

Code and configuration objects of visible pages are indexed as page content

 
View issue   ·   Add comment
 

Issue created

 
cid:jira-generated-image-avatar-d85346b9-336a-4ce1-a5e1-41a6aea48c9e Gunter Ohrner created this issue on 15/Nov/26 14:22
 
Summary: Code and configuration objects of visible pages are indexed as page content
Issue Type: cid:jira-generated-image-avatar-15fc8600-9703-4abe-a01e-4958ac1f585d Improvement
Assignee: Unassigned
Created: 15/Nov/26 14:22
Priority: cid:jira-generated-image-static-minor-52edc042-8d57-4e66-aa5c-712da8435679 Minor
Reporter: Gunter Ohrner
Description:

Collections using the XWiki document store index the objects of a page as "Metadata of Type …" sections. Hidden pages are excluded, but pages that are not hidden carry technical objects too.

In a standard wiki, a collection over Help and Sandbox indexes the CSS of a style sheet extension (Help.Videos), the parameters of a UI extension (Help.WebHome) and the live table settings of the demo applications. Other classes like this include wiki macros and components, sheet bindings, redirects, dashboard gadgets, panels, private keys and certificates, group memberships and this application's own settings.

Such text adds noise to search results and RAG context, and it is sent to the embedding and chat services. Expected: objects of classes that hold code, configuration, keys or account data are not indexed, while application data, comments, tags and user profiles still are.

Collections indexed before need a re-index to drop such chunks.