It is generally advisable to announce potential conflicts of interest while touting a particular solution in every other comment.
Agree, should mention that I'm affiliated.
The github repo has > 3k stars so does the person still have an incentive to plug their solution everywhere?
I am calculating vector embeddings on 60 sec intervals with 30 sec overlap window. So overall, I only get 20-30 embeddings for a video. So a linear lookup works fine.
to be clear: you're calculating embeddings based on the timestamped transcript, yes? or are you doing something with the actual video content? what are you using for the embeddings?
Got it. However, depends on the number of videos in the end, right? :)
I am semantically searching within 1 video only.
Hm, a search across all videos or at least a series of videos could be a valuable feature.