Are you ready to begin embedding smart blocks - each time I open Obsidian

[nomadphase]

Hi Brian,

I am being prompted to begin embedding smart blocks each time I open Obsidian. My understanding was that this would be done once when upgrading to V2 (and maybe also for a couple of additional updates since). Is it typical that it should be happening each time I open Obsidian (typically once per day).

Cheers,

[arminta7]

I am actually getting this every time I switch notes and it locks up Obsidian for about 5 secs or so.

[brianpetro]

@arminta7 I noticed this issue, too.

I started looking into it earlier today, but I still need to fix it.

I'm going to bump the priority and I'll let you know when it's fixed.

🌴

[brianpetro]

@arminta7 I found an error in the logic that I am pretty sure was causing this issue.

Instead of just embedding the current note if it wasn't already embedded, embedding all un-embedded entities was triggered.

This should be fixed in v2.0.115.

Let me know if it keeps happening and I'll give it another look.

Thanks for your help,
🌴

[brianpetro]

@nomadphase that is not supposed to be happening.

I'm wondering if the embedding file isn't saving correctly.

Do you think that 12243 blocks represents all of the blocks in your vault?

Which model are you using?

🌴

[brianpetro]

@nomadphase, impressive setup!

Unfortunately, an internal feature in JavaScript (JSON.stringify) seems to break at ~500MB, regardless of the machine specs. And that's the issue I believe we're running into.

🌴

[nomadphase]

aha - so hopefully the 1536 dimension model will come in below that constraint.

Hopefully there will be a way to overcome that constraint in the longer term, as I expect people's vaults and RAG embeddings will grow over time, even if I am currently a 'power user'.

Thank you for your support!

[brianpetro]

@nomadphase it's definitely on the roadmap and it should be done sooner than later.

There are a few different ways to go about it, and I'm still deciding which will make the most sense.

I know it's a frustrating issue, so thanks for bearing with me until it gets fixed!

🌴

[nomadphase]

Thanks Brian,

I can see you've been active on this issue, I saw your posts on the OpenAI forum from back in July!

I'm just keen to have the best information retrieval possible, whether it is local or hosted. It's important that it is reliable and that it outperforms other knowledge retrieval tech/tools.

All in all, Obsidian, Dataview, Tasks, Omnisearch, Properties, Smart Connections - it all gives me an amazing set of tools to work with a massive amount of information - it's game changing

[nomadphase]

Looks like the 1536 dim model is the go:

I shut down and restarted Obsidian and was prompted to embed 3 smart blocks! win

[destiny911]

I have been getting the embedding message very often recently and looked at my settings and realized I don't understand any of this yet.
Thought it might be helpful to ask if I should change these settings?

[brianpetro]

Hi @destiny911

No, that looks good.

The only thing you might try is setting blocks to "none" and notes to the BGE-micro model, or maybe try an OpenAI model.

But I think the issue is primarily two things that I'm in the process of addressing:

1. some logic errors in how "detecting changes notes" notes is implemented.
2. excess notifications in general

Both of these issues will be addressed in the next update. I already made some tweaks to the logic, and added the ability to "mute" notifications.

I just need to do a little more testing before I release the update.

🌴

[Derducken]

Hello @brianpetro , any news on a fix for the issue? My Ryzen 5900x with an RTX 3070 and a local LLM goes at around 700 tokens per second, and in a vault with around 3000 long-ish notes, the process seems to run whenever I launch Obsidian, locking it up for much, much more than "seconds". As I'm writing this, it's been over ten minutes I'm waiting for it to re-process the same notes, and it's only went through less than 1/3rd of their total.

[brianpetro]

Hi @Derducken

The embedding speed will depend a lot on which model you are using. That is slow, but based on your hardware, I'm guessing you selected a larger model other than the default BGE-micro, which typically makes it through my ~3000 notes in < 10 minutes.

The other thing to note is that the local embedding models are not GPU-optimized. This is a factor of using HuggingFace transformers.js, which currently doesn't support GPUs, but this should be available in the future and I'm watching the progress closely.

Lastly, the re-embedding issue can be followed here #496 (comment)

🌴

[Derducken]

Hello @brianpetro , thanks for that, but unfortunately I'm not using the supporter version. So, I'm stuck with having to recreate everything from scratch multiple times per day. Using a smaller model does make the process go faster, but also dramatically drops the "quality" of the suggestions (although that's subjective).

I guess I'll have to wait until your fixes and additions (like the new "per note" setting) reaches the "main" publicly available version.