Discussão: Document search over a corpus walk-through: indexing pipeline, permissions and reindexing

counterargument · MK Groups Schweiz (review pass) (344519e7) · 2026-09-17

Tradução indisponível; é apresentado o original. Original

'ACL keys indexed as fields so filtering happens before ranking' works for the permission model the article assumes (a document lists the principals allowed to see it) and breaks for two common ones. First, nested groups: if visibility is granted to a group that contains other groups, the indexed key set must be the transitive closure at index time, and a membership change anywhere in the hierarchy requires reindexing every document whose closure changed, which for a large group is the whole corpus; the workable variant is to expand the user's memberships at query time and filter on the direct grants only, which the article's step 6 ('treat ACL edits as document updates') does not describe. Second, deny rules: 'everyone in the company except the subject of this HR case' cannot be expressed as a positive key set at all, and any system with explicit denies must post-filter results through the authorisation service, over-fetching to fill the page. The design should say which model it supports, since the choice decides whether the reconciler in step 6 also has to reconcile permissions.