I stumbled upon that new use of mp4 format. Interesting.
So they use the MP4 container like a zip file and this is revolutionary? I mean, it’s a nice stunt, but the developer seems pretty serious about it. What am I misunderstanding?
Kinda but more like a db (ie you can search it) so instead of spinning a db for your ai you could have that store in that type of container …
So like indexed text files? I don’t understand what’s the benefit of encoding text into videos.
tl;dr - By using this very strange file format, you can functionally have access to the vast power of a vector database, but with the local simplicity of sqlite.
If I’m understanding this correctly: if you wanted to do a simple search for exact text strings, and that was all that you needed, then yes, you should probably use something like an sqlite database to index and query from.
However, if you are working with massively large data sets, and you need a vector database (for contextual or semantic searches) - well, that’s a next level tier of complexity. At that point, you need a vector database server.
What this thing does, however, is format your data into what they call “video” (but realistically would probably look like static if you were to actually play it in VLC). Then…
… I think it’s hooking into some similarities between vector databases and video processing, and then using the mature video processing technology to process the “video” at lightning-fast speeds. And you get all of that contextual power without relying on a cloud-based vector database server.
(To be clear, I’m doing a lot of hand-waving over the “similarities between vector databases and video processing” here - perhaps somebody with a computer science degree, or an autistic savant, can explain why this works the way that it does.)
Thank you for that, but I still don’t understand where the benefits come from.
Realistically what you wrote sounds like hogwash.
This dude is trying perf checking the solution https://github.com/Olow304/memvid/issues/51 , findings he has (so far) are interesting, like data getting malformed during processing (byte size comparison). Or that 16KB of text is 640KB of video.
Oh, and right now people are calling out the bullshit of the repo
https://github.com/janekm/retrieval_comparison/blob/main/memvid_critique.md
but there’s sqlite for that. how is this more?
Once again proving the addage that you can do literally fucking anything with ffmpeg
Wow. That’s impressive.