Refactor Knowledge: Lazy Loading on Show & Backend-Driven Search

[Elmolesto]

Hey everyone 👋

I've been using OpenWebUI with some very large knowledge collections (1000+ files), and I've hit some serious performance issues—mainly due to how everything is currently loaded.

The Problem

• Opening the Knowledge section downloads the entire collection, which in my case means ~70MB of data on every open.
• Triggering a command search also fetches a huge JSON with all file metadata, even before narrowing things down.

This gets pretty painful when you're dealing with big collections and remote setups—it slows down the UI and eats bandwidth unnecessarily.

What I’m Proposing

This could be improved with two main changes:

1. Lazy loading of file content in the Knowledge UI

   • Only fetch file details/content when they're actually viewed or expanded.
   • Initial load could just show basic metadata or file names.

2. Move file/collection search to the backend

   • Instead of downloading everything and filtering client-side, search logic should run server-side and return just what’s relevant.
   • This would drastically reduce the payload size and improve performance.

Why this matters

• Keeps the app snappy and responsive, even with large knowledge bases
• Reduces memory and network overhead

---

I'm happy to work on this —but wanted to bring it up here first to get your thoughts:

• Anyone else running into this with large knowledge bases?
• Would this require any major changes in backend API structure?
• Anything else to consider before starting?

Looking forward to hearing your thoughts!

[almajo]

Would love to see this! I'm also running into heavy loading times, I have at the moment 3 knowledge bases with >1500 Documents each and it slows down the experience so much. I support your idea of lazy loading by only sending/processing what the user actually needs.

Thanks for searching for the root causes and hope you find the time for a PR!

[Elmolesto]

I'll try to start the PR next week :)

[Elmolesto]

#12997 - PR for lazy loading. You're welcome to review it 🥳

[sir3mat]

Hi, it's helpful to hear you're also dealing with large knowledge bases. To effectively perform RAG on so many documents, what settings or approaches have you found most useful? Any tips or experiences you could share would be great