Did you just turn off all transcoding or pre-transcoded to 720p mp4 (or similar)?
Or did you just rely on the native apps for direct play?
Paperless, I guess OCR is very taxing, but I mainly had RAM problems regarding this on my Pi 3, with only one doc at a time, it went slow but steady.
Immich maxed out my i5 (all cores on 100) for some minutes when running the ML on my imported test library of ca. 3k Files. So yeah, when choosing the Pi4, I'd go with ente for sure!