feat(server, web): smart search filtering and pagination

[mertalev]

Description

This PR uses the new pgvecto.rs VBASE feature to implement paginated smart search. It updates the web app to take advantage of this for infinite scroll. Additionally, it refactors the custom search builder to be more reusable, integrating it into smart search. This paves the way for filtered hybrid search in the future.

Summary of changes:

• Rename smart info repository to search repository
• Move asset search method to search repository
• Make smart search repository method paginated and share the same builder (with some exclusions)
• Update web to query for the next page after reaching the bottom of the page (specifically when the second-to-last row is in view)
• Modularize typing for search builder
• Add new smart search DTO that includes most of the filters for the search builder
• Split current /search endpoint into /search/metadata and /search/smart
  • There's enough difference in what the two endpoints can accept that separating them makes validation etc. simpler
• Mark current search endpoint as deprecated
  • This is because the new search filtering logic isn't ready yet, and because making all of the necessary changes to actually switch everything to the new endpoints increases the scope of what is already a very large PR

How Has This Been Tested?

Tested with both pgvecto.rs and pgvector. The latter can still somewhat benefit from this change, but with a maximum of 1000 assets that will effectively be lower in the case of filters or multiple users.

https://github.com/immich-app/immich/assets/101130780/d2b71a06-b5c6-4104-a1b4-4c1eaa91b9b6

[cloudflare-workers-and-pages[bot]]

Deploying with    Cloudflare Pages

Latest commit:	bf879db
Status:	✅  Deploy successful!
Preview URL:	https://4f23e17d.immich.pages.dev
Branch Preview URL:	https://feat-server-web-search-pagin.immich.pages.dev

View logs

[alextran1502]

Thank you for the high quality PR.

Can you explain briefly how you achieve the infinite scrolling, especially on the performance aspect? Will it cause a performance issue if you keep scrolling since the DOM will continue to add more div as it is scrolled?

[mertalev]

Thank you for the high quality PR.

Can you explain briefly how you achieve the infinite scrolling, especially on the performance aspect? Will it cause a performance issue if you keep scrolling since the DOM will continue to add more div as it is scrolled?

Existing assets aren't re-rendered when a new page of assets are added, so I don't think this slows things down much besides likely increasing RAM usage. We also only append to the list of assets - this avoids overhead of copying or traversing this list as it gets longer.

There's definitely a point where it will get slower, not just on the frontend but also because the vector search query will have to skip past more and more rows for each page. But in an earlier version, I scrolled to about 50-60 pages deep (i.e. 5-6k assets) and didn't notice a degradation, so I'm not sure what the threshold is where it would start getting slower. It might be different for a client with less RAM, on a different browser (I tested on Chrome), etc.

[mertalev]

I can confirm it starts degrading for me at 7k+ assets. The scrolling starts getting jittery, the next page takes longer to load and the thumbnails don't show up as quickly. Continuously scrolling without ever stopping, RAM usage climbed to 1.5gb at 10k assets. But after leaving it alone to type this, it dropped down to around 650mb. Lastly, changing the width of the page changes the size of the thumbnails, meaning they have to get re-rendered. This absolutely chokes the page and results in over 3gb RAM usage with 10k assets.

[mertalev]

I think we can set a 50 page limit to prevent the number of assets from getting out of hand.