Ideas for improving video search?

carl

ak42 I think that new AI channels and AI videos are automatically created on YouTube much faster than any crowd of real humans can keep up.

ak42

haha, didn't realize it was THAT bad.
my usage of freetube is very limited to my "subscription" list and "related" videos.

on a side note: https://blog.youtube/inside-youtube/our-approach-to-responsible-ai-innovation/

But as @RoxyRoxyRoxy already said:

YouTube doesn’t really have the same incentive alignment that Kagi does

orb

Starting to see results flood with clear AI content.

See

90 second videos seem to be the target for these, would love to see reduction....

i wonder if there's a way to incorporate YouTube's captions or other metadata to help weight these results and/or red flag.

Feel free to edit/revise as needed, trying to provide as much context.

useful videos about the PS Vita.

orb

five of the videos in this screengrab are nearly identical scripts and content from channels purely contributing to dead internet.

it should be our (kagi community, users + dev) duty to try and squash these results, even if introducing "community flags" for "channels" or entire domains, enhancing the function of https://kagi.com/stats?stat=leaderboard

carl

deleted

carl

I apologize for pasting an AI generated suggestion here, but I think it could be on the right path for how to identify the most common AI generated videos, which just have still images panning and zooming while a robot voice is talking:

🛠️ Efficient Detection Strategies (Low Resource)

✅ 1. Use Keyframe Extraction Heuristics
• Extract a handful of keyframes from the video (e.g. 1 every 5–10 seconds)
• Use SSIM (Structural Similarity Index) or image hash comparison to check how similar those frames are.
• If all or most keyframes are >95% identical, it’s likely a still-image-based video.

This is very cheap computationally and can run in seconds.

✅ 2. Analyze the Video’s Motion Vectors (without full decoding)
• Many video formats include motion vector data (used by codecs like H.264 to compress video).
• If the motion vectors only represent pan or zoom and are extremely linear/static, it’s a strong signal that it’s not real video footage.

Libraries like FFmpeg can extract motion vector info without rendering full frames.

✅ 3. Check Video Metadata (FFprobe)
• Check things like:
• Frame rate vs bitrate (low entropy often = still images)
• No audio variation (flat waveform = synthetic voice or silence)
• Short duration with large resolution (often AI image slideshow)

Again, FFmpeg/FFprobe can do this very quickly.

Seeing as Kagi already are getting various frames and short clips when you hover over a thumbnail, it seems you have access to this.

RoxyRoxyRoxy

carl I can't speak really for the rest of it, but I'm pretty sure the frames you see on hover are from a YouTube API, I don't believe Kagi is generating those themselves. In that case, considering actions YouTube's been taking recently (blocking VPNs, requiring login, blocking Invidious), getting video streams at scale might not be so trivial unfortunately.

I'm going off this StackOverflow post, which has some interesting info about the frames in question. Maybe some of the techniques you brought up could be applied with the storyboard images? Not as good as the full video but I'd imagine immensely cheaper.

carl

@RoxyRoxyRoxy If those frames and short clips seen on hover is rendered by the users browser directly connecting to YouTube, then maybe the processing could be done by Kagi within the users browser? Maybe even as a "Refine" button after the initial results have loaded, in order to not make things too strange.

Maybe Kagi could get full access by paying YouTube for this like how they pay for other search engine API access. If that is costly, I would be prepared to pay extra for this as a Kagi add-on.

Crotaro

I'd love to have the option of hiding / downranking videos with ALL CAPS or emojis in their title.

Frankly, easier/different readability aside, I don't understand the current option to turn video titles into regular case or complete lowercase. When I first saw it, I thought it would automatically filter out videos that aren't my selected case structure, because that would have an effect on the actual search results.

Boomkop3

The same filter that is super nice in kagi image search, I think would work great on video search. A lot of ai generated video slop has an ai thumbnail as well. Being able to filter these would be nice

To have a lot less content-farm kind of content in the video search, optionally

« Previous Page