Repository navigation
[RFC] Files Recommendations: Improve algorithm and data sources #888
Description
Goals
I think we can improve the current state.
Right now we support three providers:
- Recently Commented
- Recently Edited
- Recently Shared
Sources suggestions
- Recently Accessed
- Recently Uploaded (because we keep the mtime when uploading, so new files might not appear as Recently Edited)
- Frequently Accessed
- Time of day: Accessed files during specific time of days ?
Aggregation bonus
- Some data could be regrouped too. Like if you frequently access more than x files in the same folder, also recommend the folder directly?
Distribution
I also suggest to add a weight combining score
Something that can be fined tuned over time, but we could weight the relevancy of a result if it matches multiple criterias (dumb values used for this example below)
$score = (
0.3 * $recencyScore +
0.2 * $frequencyScore +
0.1 * $editScore +
0.1 * $sharedScore +
0.1 * $commentedScore
);
Reacted by Marco
Activity
provokateurin commentedon Jun 26, 2025
There was already #856 which lists a bunch of issues with ideas for improvements.
[-][RFC] Improve algorithm and data sources[/-][+][RFC] Files Recommendations: Improve algorithm and data sources[/+]jancborchardt commentedon Jul 30, 2025
The biggest issue I have with recommendations right now is that the files are usually not relevant for me at all. E.g. if you check your recommended files on our company instance, it is likely to be marketing files.
What would be nice if we can consider or prefer things like:
This would help to limit the amount of files we are looking at, and thus help show you more relevant ones.
jancborchardt commentedon Jul 30, 2025
FYI @kra-mo if you have further ideas regarding this. :)
skjnldsv commentedon Jul 30, 2025
I was currently working on
Frequently Accessed.I guess adding the favorite would not be that hard
And if we do the last point of my post, it should help us adjust which source is most important :)
Frequently Accessedjancborchardt commentedon Sep 3, 2025
@skjnldsv any news regarding this by the way, or something that could need review, or where you need input? :)
skjnldsv commentedon Sep 3, 2025
It's been postponed and I'm on vacations :)
theking2 commentedon Apr 29, 2026
how about a magic file like
.ignorein a folder that should be ignored by the algorithm.Freundschaft commentedon Oct 29, 2026
ive built a plugin which just shows me the last files ive opened and its already quite useful: https://github.com/Freundschaft/nextcloud-last-opened-plugin
skjnldsv commentedon Oct 29, 2026
Picking this up again. Here's the plan, roughly 2–2.5 weeks in total.
The tricky source is "Recently/Frequently Accessed".
BeforeNodeReadEventfires for sync clients, previews and thumbnails, so counting reads would mostly measure what the desktop client touched. Tuning the weights can't fix that, the noise and the real opens end up in the same number.So I'd count actual opens instead:
NodeOpenedEventany app can listen to. It's new API, so it gets its own PR first.open(). Prev/next browsing doesn't count, so swiping through a gallery isn't 200 opens. Files, Talk, Deck, Text and Office all go through the viewer.Then favorites, recently uploaded (via
upload_time), and a weighted score that replaces the plain timestamp sort, so a file that matches several sources ranks higher.Recommending whole folders, files your colleagues use, and an ignore marker are separate follow-ups. The opens source only works on 36+.