6 comments

[ 0.21 ms ] story [ 20.1 ms ] thread
Beware 429s if you hit 100+ videos a day or so in my experience.
I was pleasantly surprised that this didn't overcomplicate the process of piping captions from yt-dlp to llm. It doesn't look like it relies on captions being available from YouTube, but this could be pretty easily overcome by adding a whisper.cpp step.
(comment deleted)
This setup is quite similar to the one fabric [1] patterns, the limitation i encountered while testing those with local llms was the prompt efficacy. More specifically the output format of the prompt is rarely respected properly.

In addition to this the tone/sentiment of the answers vary a lot between models, as usual.

Are there more compliant models wrt respecting prompt instructions? (assuming comparable parameter sizing)

[1] https://github.com/danielmiessler/fabric